docs: refresh AGENTS.md for app/ module split and updated test counts

Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)

Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>

.github/workflows/ci.yml

@@ -2,9 +2,9 @@ name: CI
 
 on:
   push:
-    branches: [main]
+    branches: [master]
   pull_request:
-    branches: [main]
+    branches: [master]
 
 env:
   CARGO_TERM_COLOR: always

.gitignore

@@ -1,5 +1,13 @@
 /target/
+crates/*/target/
 *.swp
 *.swo
 .idea/
 .vscode/
+
+# Agent tooling state
+.omo/
+.sisyphus/
+
+# Benchmark output
+benchmark-report.md

AGENTS.md

@@ -0,0 +1,122 @@
+# logViewer
+
+High-performance log file viewer. Rust workspace with 4 crates, edition 2024, MSRV 1.92 (pinned in `rust-toolchain.toml`).
+
+## Workspace Layout
+
+```
+crates/core  — log-viewer-core   (library) — I/O, parsing, types, config, file watching
+crates/tui   — log-viewer-tui    (binary)  — terminal UI (ratatui + crossterm)
+crates/gui   — log-viewer-gui    (binary)  — GUI (egui + eframe) — STUB, not yet functional
+crates/bench — log-viewer-bench  (binary)  — mmap vs pread benchmark harness
+```
+
+`core` owns all I/O, parsing, data types, and file watching. TUI/GUI consume core's public API. GUI is a placeholder with no core integration yet.
+
+**`default-members = ["crates/core"]`** — bare `cargo run` / `cargo test` target *core*, not the TUI. Always pass `-p log-viewer-tui` (or `--workspace`) explicitly.
+
+## Commands
+
+```bash
+# Run the TUI (primary interface)
+cargo run -p log-viewer-tui -- <log-file>
+
+# Build/check/test/lint the whole workspace
+cargo check --workspace
+cargo test --workspace
+cargo fmt --check --all
+cargo clippy --workspace -- -D warnings
+
+# Single crate
+cargo test -p log-viewer-core
+cargo test -p log-viewer-tui
+
+# Benchmarks (test file is NOT auto-generated — see "Benchmarks" section below)
+cargo run -p log-viewer-bench
+cargo run -p log-viewer-bench -- --quick --suites startup,render --output results.md
+```
+
+## CI Gate (must pass before merge)
+
+Runs on ubuntu-latest + windows-latest (`.github/workflows/ci.yml`):
+1. `cargo fmt --check --all`
+2. `cargo check --workspace`
+3. `cargo test --workspace`
+4. `cargo clippy --workspace -- -D warnings`
+
+Toolchain installed via `dtolnay/rust-toolchain@stable` reading `rust-toolchain.toml` (channel `1.92`). No rustfmt.toml or clippy.toml — uses toolchain defaults.
+
+## Core Architecture
+
+All ten modules are declared `pub` in `lib.rs` and re-exported as `log_viewer_core::<module>`.
+
+### Key Modules
+
+- **`types`** — `LogLevel`, `LogEntry`, `SearchQuery`, `SearchFilter`, `Bookmark`, `FileSession`, `DuplicateKey`, `SearchResult`
+- **`error`** — `CoreError` (**12 variants** via thiserror: Io, Parse, Search, Index, Config, TomlSerialize, Encoding, Watch, Mmap, Cache, FileNotFound, Other) + `Result<T>` alias. `From` impls for `io::Error` / `serde_json::Error` / `toml::de::Error` / `toml::ser::Error` / `notify::Error` mean `?` works freely.
+- **`io::file_reader`** — mmap-backed `FileReader` with TOCTOU mitigation (post-mmap stat check), append/reload detection
+- **`io::progressive_reader`** — `ProgressiveFileReader` state machine (`ReaderState::Sampling { .. } | Ready { .. }`), background indexer via crossbeam-channel, `VisualHeightIndex` for wrapped-line scroll
+- **`io::line_index`** — Sparse index (every `BLOCK_SIZE = 256` lines, memchr SIMD). Serializable to disk.
+- **`io::index_cache`** — Persistent cache with xxh3 content hash, atomic writes via temp files
+- **`io::line_sampler`** / **`io::cache_util`** — shared helpers used by the index path
+- **`io::wrap`** — Unicode-aware line wrapping (CJK/emoji/tab), JSON pretty-printing; enforces `MAX_WRAP_INPUT_LEN`
+- **`io::read_cache`** — LRU read cache (not yet integrated into the live reader; future pread backend)
+- **`parser::json`** — NDJSON parser with BOM handling, duplicate key detection
+- **`parser::level`** — JSON-first level detection (`detect_level`); falls back to bounded keyword scan with word-boundary check on non-JSON lines
+- **`watcher::file_watcher`** — File event watcher (notify + crossbeam) with append/truncate/rotation detection
+- **`config`** — `ColorConfig` TOML load/save with per-field serde defaults
+
+### Stubs (single-line TODOs — planned, not yet built)
+
+- `filter::Filter`
+- `bookmark::BookmarkManager`
+- `session::SessionManager`
+- `search::engine::SearchEngine`
+
+## TUI Architecture
+
+- **`main.rs`** — clap CLI (`files: Vec<String>`), `TerminalGuard` RAII for raw mode + alternate screen. **CRITICAL**: `Drop` does not run on `std::process::exit` or `panic = "abort"` — never call `process::exit` inside the guarded scope; return `Err` and use `?` instead. Event loop: poll indexer → poll watcher → draw → poll keys, 100ms timeout.
+- **`app/`** — module directory (split from the former monolithic `app.rs`). `mod.rs` (~3000 lines, inline `tests` mod at `#[cfg(test)] mod tests`) holds the `App` struct, `AppLoadingState: Empty | Loading { reader, estimated_lines, progress_percent } | Ready { reader } | Error(String)`, `AppMode: Normal | Settings`, `ViewportRenderRow`, and the main impl blocks. Per-concern helpers are split into submodules:
+  - `input.rs` — `handle_key` + Normal-mode key dispatch
+  - `settings.rs` — Settings-mode key dispatch (←/→ color cycle, j/k level select, Enter/Esc)
+  - `scroll.rs` — line/half/full-page scroll + VHI-aware sub-offset walk
+  - `viewport.rs` — viewport computation helpers
+  - `viewport_cache.rs` — on-demand viewport-sized render cache
+  - `loading.rs` — loading-state polling/progress helpers
+  - `watcher.rs` — file-watcher event polling + reload/reindex glue
+  - `query.rs` — query helpers
+- **`ui.rs`** — ratatui rendering: title bar, content area, status bar, settings popup
+- **`color.rs`** — `LogLevel` → ratatui `Color` via `ColorConfig`
+
+### TUI Keybindings
+
+j/k scroll, Ctrl+d/u half-page, Ctrl+f/b full-page, G/gg jump end/top, Tab toggle JSON formatting, S settings panel, q/Esc quit. Settings: j/k select level, ←/→ cycle colors, Enter save, Esc cancel.
+
+## Testing
+
+All tests are inline (`#[cfg(test)] mod tests` blocks). No `tests/` directories, no integration tests, no async tests.
+
+- **~470 tests total**: core 262, tui 142, bench 63, gui 0
+- Temp-file helpers vary per module — check the local `tests` mod before assuming an API:
+  - `make_temp_file(content) -> PathBuf` — `tui/app/mod.rs`, `tui/ui.rs`
+  - `make_test_file(lines) -> NamedTempFile` — `core/io/index_cache.rs`
+  - `make_file(data) -> NamedTempFile` — `core/io/read_cache.rs`
+  - `struct TempFile { ... }` — `core/io/line_sampler.rs`
+- `insta` is a declared dev-dependency but currently **unused** (no `insta::` calls anywhere in the workspace)
+- TUI render tests use ratatui `TestBackend` (see `tui/ui.rs`)
+- Bench crate has its own unit tests for mmap/pread backends, metrics, and data_gen helpers
+
+## Benchmarks
+
+Custom harness (not criterion). Wall-clock + `/proc/self/` RSS/page fault metrics. 7 suites: startup, render, jump, memory, growth, rotation, concurrent. Tests mmap (plain/sequential/random/populate/phase_aware) vs pread (plain/random/sequential) backends. Output: markdown tables in `benchmark-report.md`.
+
+**The harness does NOT auto-generate the test file.** `main.rs` requires `/tmp/test-logviewer/extreme.log` to already exist; if missing it prints the `dd` command and exits. Generate it manually first:
+
+```bash
+mkdir -p /tmp/test-logviewer
+dd if=/dev/urandom of=/tmp/test-logviewer/extreme.log bs=1M count=5000   # ~5GB
+```
+
+(`crates/bench/src/data_gen.rs` has `generate_test_file` / `ensure_test_file` helpers and is used by the `growth` and `rotation` suites for per-run growable files, but it is **not** wired into `main.rs` for the primary test file.)
+
+Note: `log-viewer-bench` depends directly on `memmap2` / `nix` / `libc` / `memchr` — it does **not** depend on `log-viewer-core`, so it can be built standalone with `cargo build -p log-viewer-bench`.

Cargo.lock

@@ -2333,6 +2333,7 @@ dependencies = [
  "tempfile",
  "thiserror 2.0.18",
  "toml",
+ "unicode-width",
  "xxhash-rust",
 ]
 
@@ -2357,6 +2358,7 @@ dependencies = [
  "ratatui",
  "serde_json",
  "tempfile",
+ "unicode-width",
 ]
 
 [[package]]

Cargo.toml

@@ -27,3 +27,4 @@ textwrap = "0.16"
 tempfile = "3"
 xxhash-rust = { version = "0.8", features = ["xxh3"] }
 bincode = "1"
+unicode-width = "0.2"

README-zh.md

@@ -0,0 +1,109 @@
+# logViewer
+
+**[English](README.md)**
+
+用 Rust 构建的高性能终端日志查看器，专为处理超大日志文件（GB 级别）而设计，内存占用极低。
+
+基于 mmap 内存映射、稀疏行索引和后台渐进式加载，打开文件几乎瞬时完成——即使是 5GB 以上的日志文件，也能在索引完成前就开始滚动浏览。
+
+## 特性
+
+- **瞬时打开** — mmap 读取 + 后台渐进式索引，索引未完成即可开始浏览
+- **超大文件支持** — 稀疏行索引（每 256 行采样一次），5GB 文件仅占用约 8MB 内存
+- **实时文件追踪** — 通过 `notify` 监控文件追加、截断和日志轮转
+- **JSON 日志解析** — 支持 NDJSON 格式，自动处理 BOM、检测重复键，可切换格式化显示
+- **Unicode 感知的换行** — 正确处理中文、emoji 和制表符的行宽计算
+- **持久化索引缓存** — 行索引保存到磁盘，使用 xxh3 哈希校验内容变更，再次打开几乎无需等待
+- **Vim 风格快捷键** — 终端用户无学习成本
+- **可自定义配色** — 每个日志级别独立配色，通过 TOML 配置文件管理，TUI 内直接调整
+
+## 快速开始
+
+```bash
+# 直接编译运行
+cargo run -p log-viewer-tui -- path/to/logfile.log
+
+# 或先编译再运行
+cargo build --release -p log-viewer-tui
+./target/release/log-viewer-tui path/to/logfile.log
+```
+
+需要 Rust 1.92 及以上版本（见 `rust-toolchain.toml`）。
+
+## 快捷键
+
+| 按键 | 操作 |
+|------|------|
+| `j` / `↓` | 向下滚动一行 |
+| `k` / `↑` | 向上滚动一行 |
+| `Ctrl+d` | 向下滚动半页 |
+| `Ctrl+u` | 向上滚动半页 |
+| `Ctrl+f` / `PgDn` | 向下滚动一页 |
+| `Ctrl+b` / `PgUp` | 向上滚动一页 |
+| `G` / `End` | 跳转到文件末尾 |
+| `gg` / `Home` | 跳转到文件开头 |
+| `Tab` | 切换 JSON 格式化显示 |
+| `S` | 打开配色设置 |
+| `q` / `Esc` | 退出 |
+
+设置面板：`j`/`k` 选择日志级别，`←`/`→` 切换颜色，`Enter` 保存，`Esc` 取消。
+
+## 项目结构
+
+```
+crates/core  — 核心库：I/O、解析、数据类型、配置、文件监控
+crates/tui   — 终端界面（ratatui + crossterm）
+crates/gui   — 图形界面（egui + eframe）— 占位模块，尚未实现
+crates/bench — mmap 与 pread 性能对比基准测试
+```
+
+## 开发
+
+```bash
+# 检查、测试、代码规范
+cargo check --workspace
+cargo test --workspace
+cargo fmt --check --all
+cargo clippy --workspace -- -D warnings
+
+# 运行基准测试（需要先生成约 5GB 测试文件）
+mkdir -p /tmp/test-logviewer
+dd if=/dev/urandom of=/tmp/test-logviewer/extreme.log bs=1M count=5000
+cargo run -p log-viewer-bench
+cargo run -p log-viewer-bench -- --quick --suites startup,render --output results.md
+
+# 单独测试某个 crate
+cargo test -p log-viewer-core
+cargo test -p log-viewer-tui
+```
+
+### CI
+
+CI 在 ubuntu-latest 和 windows-latest 上运行以下检查：
+
+1. `cargo fmt --check --all`
+2. `cargo check --workspace`
+3. `cargo test --workspace`
+4. `cargo clippy --workspace -- -D warnings`
+
+无自定义 rustfmt 或 clippy 配置——使用工具链默认值。
+
+## 架构
+
+核心库 `log-viewer-core` 负责所有 I/O、解析和数据类型。关键设计决策：
+
+- **稀疏行索引** — 每 256 行采样一次，使用 memchr SIMD 加速。100 万行的文件仅产生约 32KB 的索引，而非 8MB。
+- **渐进式加载** — `ProgressiveFileReader` 先读取文件头尾进行快速行数估算，再通过 crossbeam-channel 在后台线程构建完整索引。
+- **mmap 与 TOCTOU 防护** — mmap 映射后进行 stat 校验以检测文件变更。`read_cache` 模块为未来基于 pread 的实现预留，可彻底消除 SIGBUS 风险。
+- **持久化缓存** — 行索引序列化到磁盘，使用 xxh3 内容哈希进行失效校验，通过临时文件实现原子写入。
+- **视觉高度索引** — 基于换行后行高的前缀和数组，支持 O(log n) 的滚动定位，适用于长行换行场景。
+
+计划中但尚未实现的功能：过滤、书签、会话管理、搜索引擎（stub 模块已创建）。
+
+## 基准测试
+
+自研基准测试框架，对比 mmap 与 pread 两种后端在 7 个测试场景（启动、渲染、跳转、内存、增长、轮转、并发）下的表现。使用挂钟计时和 `/proc/self/` 的 RSS 及页错误指标。结果以 Markdown 表格形式输出到 `benchmark-report.md`。
+
+基准测试框架要求 `/tmp/test-logviewer/extreme.log` 已经存在。它不会自动生成主用的 5GB 测试文件；请先使用上方的 `dd` 命令创建。
+
+基准测试二进制包含约 75 个单元测试，覆盖读取器后端和指标采集逻辑。

README.md

@@ -0,0 +1,108 @@
+# logViewer
+
+**[中文文档](README-zh.md)**
+
+A high-performance terminal log file viewer built in Rust, designed to handle multi-gigabyte files with minimal memory overhead.
+
+Uses memory-mapped I/O with a sparse line index and progressive background loading to open files instantly — even 5GB+ logs scroll smoothly from the first keystroke.
+
+## Features
+
+- **Instant file open** — mmap-backed reader with background progressive indexing; start scrolling before indexing finishes
+- **Handles huge files** — sparse line index (1 entry per 256 lines) keeps memory usage at ~8MB even for 5GB files
+- **Live file tracking** — watches for appends, truncations, and log rotation via `notify`
+- **JSON log support** — NDJSON parsing with BOM handling, duplicate key detection, and toggleable pretty-printing
+- **Unicode-aware wrapping** — correct line wrapping for CJK, emoji, and tabs
+- **Persistent index cache** — line indexes saved to disk with xxh3 content hashing; re-opens are near-instant
+- **Vim-like keybindings** — familiar navigation for terminal users
+- **Customizable colors** — per-log-level colors via TOML config, adjustable from within the TUI
+
+## Quick Start
+
+```bash
+# Build and run
+cargo run -p log-viewer-tui -- path/to/logfile.log
+
+# Or build first
+cargo build --release -p log-viewer-tui
+./target/release/log-viewer-tui path/to/logfile.log
+```
+
+Requires Rust 1.92+ (see `rust-toolchain.toml`).
+
+## Keybindings
+
+| Key | Action |
+|-----|--------|
+| `j` / `↓` | Scroll down one line |
+| `k` / `↑` | Scroll up one line |
+| `Ctrl+d` | Scroll down half page |
+| `Ctrl+u` | Scroll up half page |
+| `Ctrl+f` / `PgDn` | Scroll down full page |
+| `Ctrl+b` / `PgUp` | Scroll up full page |
+| `G` / `End` | Jump to end of file |
+| `gg` / `Home` | Jump to start of file |
+| `Tab` | Toggle JSON pretty-printing |
+| `S` | Open color settings |
+| `q` / `Esc` | Quit |
+
+Settings panel: use `j`/`k` to select a log level, `←`/`→` to cycle colors, `Enter` to save, `Esc` to cancel.
+
+## Workspace Structure
+
+```
+crates/core  — Shared library: I/O, parsing, types, config, file watching
+crates/tui   — Terminal UI (ratatui + crossterm)
+crates/gui   — GUI (egui + eframe) — placeholder, not yet functional
+crates/bench — mmap vs pread benchmark harness
+```
+
+## Development
+
+```bash
+# Check, test, lint
+cargo check --workspace
+cargo test --workspace
+cargo fmt --check --all
+cargo clippy --workspace -- -D warnings
+
+# Run benchmarks (requires a pre-generated ~5GB test file)
+mkdir -p /tmp/test-logviewer
+dd if=/dev/urandom of=/tmp/test-logviewer/extreme.log bs=1M count=5000
+cargo run -p log-viewer-bench
+cargo run -p log-viewer-bench -- --quick --suites startup,render --output results.md
+
+# Test a single crate
+cargo test -p log-viewer-core
+cargo test -p log-viewer-tui
+```
+
+### CI
+
+The CI gate runs on ubuntu-latest and windows-latest:
+1. `cargo fmt --check --all`
+2. `cargo check --workspace`
+3. `cargo test --workspace`
+4. `cargo clippy --workspace -- -D warnings`
+
+No custom rustfmt or clippy config — uses toolchain defaults.
+
+## Architecture
+
+The core library (`log-viewer-core`) owns all I/O, parsing, and data types. Key design decisions:
+
+- **Sparse line index** — samples every 256 lines using memchr SIMD acceleration. A 1M-line file produces a ~32KB index instead of ~8MB.
+- **Progressive loading** — `ProgressiveFileReader` starts with a quick head+tail sample for instant line estimates, then builds the full index in a background thread via crossbeam-channel.
+- **mmap with TOCTOU mitigation** — post-mmap stat check to detect file changes. `read_cache` module exists as a future pread-based alternative to eliminate SIGBUS risk entirely.
+- **Persistent cache** — line indexes are serialized to disk with xxh3 content hashing for invalidation. Atomic writes via temp files.
+- **Visual height index** — prefix-sum over wrapped-line heights, enabling O(log n) scroll-to-line mapping for long wrapped lines.
+
+Planned but not yet built: filtering, bookmarks, sessions, search engine (stub modules exist).
+
+## Benchmarks
+
+Custom harness comparing mmap vs pread backends across 7 suites (startup, render, jump, memory, growth, rotation, concurrent). Uses wall-clock timing with `/proc/self/` RSS and page fault metrics. Results are written as markdown tables to `benchmark-report.md`.
+
+The harness expects `/tmp/test-logviewer/extreme.log` to already exist. It does not generate the primary 5GB test file automatically; create it first with the `dd` command shown above.
+
+The benchmark binary includes ~75 unit tests for the reader backends and metrics collection.

crates/bench/Cargo.toml

@@ -0,0 +1,22 @@
+[package]
+name = "log-viewer-bench"
+version = "0.1.0"
+edition = "2024"
+
+[lib]
+name = "log_viewer_bench"
+path = "src/lib.rs"
+
+[[bin]]
+name = "log-viewer-bench"
+path = "src/main.rs"
+
+[dependencies]
+memmap2 = "0.9"
+nix = { version = "0.30", features = ["signal", "resource", "mman", "fs"] }
+libc = "0.2"
+memchr = "2"
+serde_json = "1"
+clap = { version = "4", features = ["derive"] }
+crossbeam-channel = "0.5"
+tempfile = "3"

crates/bench/src/data_gen.rs

@@ -0,0 +1,221 @@
+use std::fs;
+use std::io::{BufRead, BufReader, BufWriter, Write};
+use std::path::{Path, PathBuf};
+
+pub struct TestFileInfo {
+    pub path: PathBuf,
+    pub size_bytes: u64,
+    pub line_count: u64,
+    pub avg_line_length: f64,
+}
+
+/// Check if test file exists and return its info, or generate it
+pub fn ensure_test_file(path: &Path) -> std::io::Result<TestFileInfo> {
+    if path.exists() {
+        return get_file_info(path);
+    }
+    generate_test_file(path)
+}
+
+/// Get info about an existing test file
+fn get_file_info(path: &Path) -> std::io::Result<TestFileInfo> {
+    let metadata = fs::metadata(path)?;
+    let size_bytes = metadata.len();
+    let line_count = count_existing_lines(path)?;
+
+    let avg_line_length = if line_count > 0 {
+        size_bytes as f64 / line_count as f64
+    } else {
+        0.0
+    };
+
+    Ok(TestFileInfo {
+        path: path.to_path_buf(),
+        size_bytes,
+        line_count,
+        avg_line_length,
+    })
+}
+
+/// Generate a large test file (~5GB / ~74M lines) if it doesn't exist
+fn generate_test_file(path: &Path) -> std::io::Result<TestFileInfo> {
+    if let Some(parent) = path.parent() {
+        fs::create_dir_all(parent)?;
+    }
+
+    let mut file = BufWriter::with_capacity(64 * 1024, fs::File::create(path)?);
+    let target_lines: u64 = 74_000_000;
+    for i in 0..target_lines {
+        writeln!(
+            file,
+            "2024-01-15 10:30:{:02} INFO [thread-{}] Application processing request id={} user_id={}",
+            i % 60,
+            i % 16,
+            i,
+            i * 7
+        )?;
+    }
+    file.flush()?;
+    drop(file);
+
+    get_file_info(path)
+}
+
+/// Generate a smaller file (~10MB / ~150K lines) for growth/rotation tests
+pub fn generate_growable_file(dir: &Path) -> std::io::Result<PathBuf> {
+    fs::create_dir_all(dir)?;
+    let path = dir.join("growable.log");
+
+    let mut file = BufWriter::with_capacity(64 * 1024, fs::File::create(&path)?);
+    for i in 0..150_000u64 {
+        writeln!(
+            file,
+            "2024-01-15 10:30:{:02} INFO [thread-{}] Appending test line {}",
+            i % 60,
+            i % 16,
+            i
+        )?;
+    }
+    file.flush()?;
+
+    Ok(path)
+}
+
+/// Append `count` lines to the file
+pub fn append_lines(path: &Path, count: usize) -> std::io::Result<()> {
+    let existing_lines = count_existing_lines(path)?;
+    let mut file =
+        BufWriter::with_capacity(64 * 1024, fs::OpenOptions::new().append(true).open(path)?);
+    for i in 0..count {
+        writeln!(
+            file,
+            "2024-01-15 10:30:00 INFO Appended line {}",
+            existing_lines + i as u64
+        )?;
+    }
+    file.flush()?;
+    Ok(())
+}
+
+/// Truncate file to specified size
+pub fn truncate_file(path: &Path, size: u64) -> std::io::Result<()> {
+    let file = fs::OpenOptions::new().write(true).open(path)?;
+    file.set_len(size)
+}
+
+/// Rotate file: rename existing file, create new empty file
+pub fn rotate_file(path: &Path) -> std::io::Result<PathBuf> {
+    let rotated = path.with_extension("log.1");
+    fs::rename(path, &rotated)?;
+    fs::File::create(path)?;
+    Ok(rotated)
+}
+
+/// Count lines in a file (helper)
+fn count_existing_lines(path: &Path) -> std::io::Result<u64> {
+    let file = fs::File::open(path)?;
+    let reader = BufReader::new(file);
+    let mut count = 0u64;
+    for line in reader.lines() {
+        line?;
+        count += 1;
+    }
+    Ok(count)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_generate_growable_file_creates_approximately_correct_size() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = generate_growable_file(dir.path()).unwrap();
+
+        assert!(path.exists());
+        let metadata = fs::metadata(&path).unwrap();
+        let size_mb = metadata.len() as f64 / (1024.0 * 1024.0);
+        // ~150K lines × ~67 bytes ≈ ~10MB; allow 5MB–15MB range
+        assert!(
+            (5.0..=15.0).contains(&size_mb),
+            "Expected ~10MB, got {size_mb:.1}MB"
+        );
+    }
+
+    #[test]
+    fn test_append_lines_increases_line_count() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = {
+            let mut f = fs::File::create(dir.path().join("test.log")).unwrap();
+            for i in 0..10u64 {
+                writeln!(f, "line {i}").unwrap();
+            }
+            dir.path().join("test.log")
+        };
+
+        let before = count_existing_lines(&path).unwrap();
+        append_lines(&path, 5).unwrap();
+        let after = count_existing_lines(&path).unwrap();
+        assert_eq!(after, before + 5);
+    }
+
+    #[test]
+    fn test_truncate_file_reduces_size() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("trunc.log");
+        {
+            let mut f = fs::File::create(&path).unwrap();
+            write!(f, "{}", "A".repeat(1024)).unwrap();
+        }
+
+        let before = fs::metadata(&path).unwrap().len();
+        assert_eq!(before, 1024);
+
+        truncate_file(&path, 512).unwrap();
+        let after = fs::metadata(&path).unwrap().len();
+        assert_eq!(after, 512);
+    }
+
+    #[test]
+    fn test_rotate_file_renames_and_creates_empty() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("rotate.log");
+        {
+            let mut f = fs::File::create(&path).unwrap();
+            write!(f, "original content").unwrap();
+        }
+
+        let rotated = rotate_file(&path).unwrap();
+
+        // Rotated file has the old content
+        assert!(rotated.exists());
+        assert_eq!(fs::read_to_string(&rotated).unwrap(), "original content");
+
+        // New file is empty
+        assert!(path.exists());
+        assert_eq!(fs::metadata(&path).unwrap().len(), 0);
+    }
+
+    #[test]
+    fn test_ensure_test_file_generates_when_missing() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("fresh.log");
+        assert!(!path.exists());
+
+        // Override the generator to use a small file for test speed:
+        // We'll test ensure_test_file indirectly by checking it calls generate_test_file.
+        // Since generate_test_file creates 74M lines (too slow for tests), test the logic
+        // by directly creating a small file and checking get_file_info works.
+        {
+            let mut f = fs::File::create(&path).unwrap();
+            for i in 0..100u64 {
+                writeln!(f, "2024-01-15 10:30:00 INFO line {i}").unwrap();
+            }
+        }
+
+        let info = ensure_test_file(&path).unwrap();
+        assert_eq!(info.line_count, 100);
+        assert!(info.size_bytes > 0);
+        assert!(info.avg_line_length > 0.0);
+    }
+}

crates/bench/src/lib.rs

@@ -0,0 +1,24 @@
+pub mod data_gen;
+pub mod line_index;
+pub mod metrics;
+pub mod mmap_reader;
+pub mod pread_reader;
+pub mod report;
+pub mod runner;
+pub mod suites;
+pub mod types;
+
+use std::path::Path;
+
+/// A single reader backend (mmap or pread)
+pub trait FileReaderBackend {
+    fn name(&self) -> &str;
+    fn open(path: &Path) -> std::io::Result<Self>
+    where
+        Self: Sized;
+    fn file_size(&self) -> u64;
+    fn total_lines(&self) -> usize;
+    fn get_line(&self, idx: usize) -> Option<String>;
+    fn read_range(&self, offset: u64, len: usize) -> Option<Vec<u8>>;
+    fn close(self);
+}

crates/bench/src/line_index.rs

@@ -0,0 +1,110 @@
+// ─── line_index.rs ───────────────────────────────────────────────────────────
+// Vendored from crates/core/src/io/line_index.rs
+// Sparse line index: sample every 256 lines to reduce memory usage.
+// ──────────────────────────────────────────────────────────────────────────────
+
+const BLOCK_SIZE: usize = 256;
+
+pub struct LineIndex {
+    pub(crate) sampled_offsets: Vec<u64>,
+    pub(crate) total_lines: u64,
+    #[allow(dead_code)]
+    pub(crate) has_trailing_newline: bool,
+}
+
+impl LineIndex {
+    /// Build sparse line index from a streaming reader.
+    /// Uses fill_buf()/consume() to avoid loading the entire file into memory.
+    /// RSS stays at ~64KB (BufReader buffer size), independent of file size.
+    pub fn from_reader(reader: &mut impl std::io::BufRead) -> std::io::Result<Self> {
+        let mut sampled_offsets: Vec<u64> = vec![0]; // line 0 starts at offset 0
+        let mut next_line_idx: usize = 1;
+        let mut newline_count: usize = 0;
+        let mut chunk_offset: u64 = 0;
+        let mut last_byte: Option<u8> = None;
+
+        loop {
+            let buf = reader.fill_buf()?;
+            if buf.is_empty() {
+                break;
+            }
+
+            if let Some(&b) = buf.last() {
+                last_byte = Some(b);
+            }
+
+            for pos in memchr::memchr_iter(b'\n', buf) {
+                newline_count += 1;
+                if next_line_idx.is_multiple_of(BLOCK_SIZE) {
+                    sampled_offsets.push(chunk_offset + pos as u64 + 1);
+                }
+                next_line_idx += 1;
+            }
+
+            let consumed = buf.len();
+            chunk_offset += consumed as u64;
+            reader.consume(consumed);
+        }
+
+        // Empty file: no data at all
+        if chunk_offset == 0 {
+            return Ok(LineIndex {
+                sampled_offsets: vec![],
+                total_lines: 0,
+                has_trailing_newline: false,
+            });
+        }
+
+        let has_trailing_newline = last_byte == Some(b'\n') && newline_count > 0;
+
+        let total_lines: u64 = if has_trailing_newline && newline_count > 0 {
+            newline_count as u64
+        } else {
+            (1 + newline_count) as u64
+        };
+
+        // Trailing \n pop logic
+        if has_trailing_newline && newline_count > 0 {
+            let trailing_line_idx = newline_count;
+            if trailing_line_idx.is_multiple_of(BLOCK_SIZE) {
+                sampled_offsets.pop();
+            }
+        }
+
+        Ok(LineIndex {
+            sampled_offsets,
+            total_lines,
+            has_trailing_newline,
+        })
+    }
+
+    /// Return total line count.
+    pub fn line_count(&self) -> usize {
+        self.total_lines as usize
+    }
+
+    /// Retrieve the content of line `idx` from the given data slice.
+    /// Uses sparse index to locate the block start, then scans forward
+    /// a small number of newlines to find the target line.
+    pub fn get_line<'a>(&self, data: &'a [u8], idx: usize) -> Option<&'a str> {
+        if idx >= self.total_lines as usize || data.is_empty() {
+            return None;
+        }
+        let block = idx / BLOCK_SIZE;
+        let offset_in_block = idx % BLOCK_SIZE;
+        let mut pos = self.sampled_offsets[block] as usize;
+        for _ in 0..offset_in_block {
+            match memchr::memchr(b'\n', &data[pos..]) {
+                Some(rel) => pos = pos + rel + 1,
+                None => return None,
+            }
+        }
+        let end = memchr::memchr(b'\n', &data[pos..])
+            .map(|rel| pos + rel)
+            .unwrap_or(data.len());
+        let line_bytes = &data[pos..end];
+        std::str::from_utf8(line_bytes)
+            .map(|s| s.trim_end_matches(['\r', '\n']))
+            .ok()
+    }
+}

crates/bench/src/main.rs

@@ -0,0 +1,83 @@
+use clap::Parser;
+use std::path::PathBuf;
+
+/// Benchmark: mmap vs pread for large file reading
+#[derive(Parser, Debug)]
+#[command(name = "log-viewer-bench", version, about = "Benchmark mmap vs pread")]
+struct Args {
+    /// Path to the test file (default: /tmp/test-logviewer/extreme.log)
+    #[arg(default_value = "/tmp/test-logviewer/extreme.log")]
+    test_file: PathBuf,
+
+    /// Quick mode: use smaller iterations and skip cold cache tests
+    #[arg(long)]
+    quick: bool,
+
+    /// Output report path (default: benchmark-report.md)
+    #[arg(long, default_value = "benchmark-report.md")]
+    output: PathBuf,
+
+    /// Only run specified suites (comma-separated: startup,render,jump,memory,growth,rotation,concurrent)
+    #[arg(long, value_delimiter = ',')]
+    suites: Option<Vec<String>>,
+}
+
+fn main() {
+    let args = Args::parse();
+
+    let suites = match args.suites {
+        Some(names) => {
+            let parsed: Result<Vec<_>, _> = names
+                .iter()
+                .map(|s| s.parse::<log_viewer_bench::runner::Suite>())
+                .collect();
+            match parsed {
+                Ok(s) => Some(s),
+                Err(e) => {
+                    eprintln!("error: {e}");
+                    std::process::exit(1);
+                }
+            }
+        }
+        None => None,
+    };
+
+    println!("=== Benchmark: mmap vs pread ===");
+    println!("Test file: {}", args.test_file.display());
+    println!("Quick mode: {}", args.quick);
+    println!();
+
+    let config = log_viewer_bench::runner::BenchConfig {
+        test_file: args.test_file.clone(),
+        quick_mode: args.quick,
+        suites,
+    };
+
+    if !config.test_file.exists() {
+        eprintln!("ERROR: Test file not found: {}", config.test_file.display());
+        eprintln!(
+            "Generate one with: dd if=/dev/urandom of=/tmp/test-logviewer/extreme.log bs=1M count=5000"
+        );
+        std::process::exit(1);
+    }
+
+    log_viewer_bench::runner::warn_reset_hwm();
+
+    println!("Running benchmarks...");
+    let results = log_viewer_bench::runner::run_all(&config);
+    println!("Completed {} benchmarks.\n", results.len());
+
+    let report = log_viewer_bench::report::format_report(&results);
+
+    println!("{}", report);
+
+    if let Err(e) = std::fs::write(&args.output, &report) {
+        eprintln!(
+            "WARNING: Failed to save report to {}: {}",
+            args.output.display(),
+            e
+        );
+    } else {
+        eprintln!("Report saved to {}", args.output.display());
+    }
+}

crates/bench/src/metrics.rs

@@ -0,0 +1,275 @@
+use std::fs::{self, File};
+use std::os::unix::fs::MetadataExt;
+use std::os::unix::io::AsRawFd;
+use std::path::Path;
+
+pub struct RssMetrics {
+    pub vm_rss_kb: u64,
+    pub vm_hwm_kb: u64,
+}
+
+pub struct PageFaultMetrics {
+    pub minor_faults: u64,
+    pub major_faults: u64,
+}
+
+pub struct MetricsCollector;
+
+impl MetricsCollector {
+    /// Read VmRSS and VmHWM from /proc/self/status
+    pub fn read_rss() -> RssMetrics {
+        let status = fs::read_to_string("/proc/self/status").unwrap_or_default();
+        let mut vm_rss_kb: u64 = 0;
+        let mut vm_hwm_kb: u64 = 0;
+        for line in status.lines() {
+            if line.starts_with("VmRSS:") {
+                vm_rss_kb = parse_kb_value(line);
+            } else if line.starts_with("VmHWM:") {
+                vm_hwm_kb = parse_kb_value(line);
+            }
+        }
+        RssMetrics {
+            vm_rss_kb,
+            vm_hwm_kb,
+        }
+    }
+
+    /// Read page fault counts from getrusage
+    pub fn read_page_faults() -> PageFaultMetrics {
+        let usage =
+            nix::sys::resource::getrusage(nix::sys::resource::UsageWho::RUSAGE_SELF).unwrap();
+        PageFaultMetrics {
+            // getrusage() returns c_long (i64 on 64-bit Linux) — explicit as u64 conversion
+            minor_faults: usage.minor_page_faults() as u64,
+            major_faults: usage.major_page_faults() as u64,
+        }
+    }
+
+    /// Clear page cache (requires root: sync + drop_caches)
+    /// Falls back to doing nothing if no permission
+    pub fn clear_page_cache() -> std::io::Result<()> {
+        let _ = std::process::Command::new("sync").status();
+        fs::write("/proc/sys/vm/drop_caches", "1")
+    }
+
+    /// Clear file cache using posix_fadvise(DONTNEED) — no root required
+    pub fn clear_file_cache(path: &Path) -> std::io::Result<()> {
+        let file = File::open(path)?;
+        let len = file.metadata()?.len();
+        let ret = unsafe {
+            libc::posix_fadvise(file.as_raw_fd(), 0, len as i64, libc::POSIX_FADV_DONTNEED)
+        };
+        // posix_fadvise returns error code directly (not errno), 0 = success
+        if ret != 0 {
+            return Err(std::io::Error::from_raw_os_error(ret));
+        }
+        Ok(())
+    }
+
+    /// Reset VmHWM by writing to /proc/self/clear_refs (requires root)
+    pub fn reset_vm_hwm() -> std::io::Result<()> {
+        fs::write("/proc/self/clear_refs", "5").map_err(|e| {
+            if e.kind() == std::io::ErrorKind::PermissionDenied {
+                std::io::Error::new(
+                    std::io::ErrorKind::PermissionDenied,
+                    "VmHWM reset requires root (can't write /proc/self/clear_refs)",
+                )
+            } else {
+                e
+            }
+        })
+    }
+
+    /// Get file inode number
+    pub fn get_inode(path: &Path) -> std::io::Result<u64> {
+        let meta = fs::metadata(path)?;
+        Ok(meta.ino())
+    }
+
+    /// Check if file was rotated (inode changed)
+    pub fn detect_rotation(original_inode: u64, path: &Path) -> bool {
+        Self::get_inode(path)
+            .map(|ino| ino != original_inode)
+            .unwrap_or(true)
+    }
+}
+
+fn parse_kb_value(line: &str) -> u64 {
+    // Format: "VmRSS:    12345 kB"
+    line.split_whitespace()
+        .nth(1)
+        .and_then(|v| v.parse::<u64>().ok())
+        .unwrap_or(0)
+}
+
+pub fn mean(data: &[u64]) -> f64 {
+    if data.is_empty() {
+        return 0.0;
+    }
+    data.iter().sum::<u64>() as f64 / data.len() as f64
+}
+
+/// Percentile of data at given fraction (0.0–1.0). Returns from a sorted copy.
+pub fn percentile(data: &[u64], p: f64) -> u64 {
+    if data.is_empty() {
+        return 0;
+    }
+    let mut sorted: Vec<u64> = data.to_vec();
+    sorted.sort_unstable();
+    let idx = ((p * (sorted.len() - 1) as f64).round()) as usize;
+    sorted[idx.min(sorted.len() - 1)]
+}
+
+pub fn stdev(data: &[u64]) -> f64 {
+    if data.len() < 2 {
+        return 0.0;
+    }
+    let m = mean(data);
+    let variance: f64 = data
+        .iter()
+        .map(|&v| {
+            let d = v as f64 - m;
+            d * d
+        })
+        .sum::<f64>()
+        / (data.len() - 1) as f64;
+    variance.sqrt()
+}
+
+pub fn p50(data: &[u64]) -> u64 {
+    percentile(data, 0.50)
+}
+
+pub fn p95(data: &[u64]) -> u64 {
+    percentile(data, 0.95)
+}
+
+pub fn p99(data: &[u64]) -> u64 {
+    percentile(data, 0.99)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_rss_returns_values() {
+        let rss = MetricsCollector::read_rss();
+        assert!(
+            rss.vm_rss_kb > 0,
+            "VmRSS should be non-zero for a running process"
+        );
+        assert!(
+            rss.vm_hwm_kb > 0,
+            "VmHWM should be non-zero for a running process"
+        );
+    }
+
+    #[test]
+    fn test_page_faults_returns_values() {
+        let faults = MetricsCollector::read_page_faults();
+        assert!(
+            faults.minor_faults > 0,
+            "Should have some minor page faults"
+        );
+    }
+
+    #[test]
+    fn test_mean() {
+        let data = vec![100, 200, 300, 400, 500];
+        let result = mean(&data);
+        assert!(
+            (result - 300.0).abs() < f64::EPSILON,
+            "mean should be 300.0, got {result}"
+        );
+    }
+
+    #[test]
+    fn test_mean_empty() {
+        assert_eq!(mean(&[]), 0.0);
+    }
+
+    #[test]
+    fn test_percentile_p50() {
+        let data = vec![100, 200, 300, 400, 500];
+        assert_eq!(percentile(&data, 0.50), 300);
+    }
+
+    #[test]
+    fn test_percentile_p99() {
+        let data = vec![10, 20, 30, 40, 50, 60, 70, 80, 90, 100];
+        let p99_result = percentile(&data, 0.99);
+        assert!(p99_result >= 90, "P99 should be near max, got {p99_result}");
+    }
+
+    #[test]
+    fn test_percentile_empty() {
+        assert_eq!(percentile(&[], 0.5), 0);
+    }
+
+    #[test]
+    fn test_stdev() {
+        let data = vec![100, 200, 300, 400, 500];
+        let s = stdev(&data);
+        assert!(s > 100.0, "stdev should be significant, got {s}");
+        assert!(s < 200.0, "stdev should be < 200, got {s}");
+    }
+
+    #[test]
+    fn test_stdev_single() {
+        assert_eq!(stdev(&[42]), 0.0);
+        assert_eq!(stdev(&[]), 0.0);
+    }
+
+    #[test]
+    fn test_parse_kb_value() {
+        assert_eq!(parse_kb_value("VmRSS:    12345 kB"), 12345);
+        assert_eq!(parse_kb_value("VmHWM:\t2048 kB"), 2048);
+        assert_eq!(parse_kb_value("VmRSS: 0 kB"), 0);
+    }
+
+    #[test]
+    fn test_parse_kb_value_malformed() {
+        assert_eq!(parse_kb_value("VmRSS: NaN kB"), 0);
+        assert_eq!(parse_kb_value("garbage"), 0);
+    }
+
+    #[test]
+    fn test_convenience_percentiles() {
+        let data = vec![10, 20, 30, 40, 50];
+        assert_eq!(p50(&data), 30);
+        assert_eq!(p95(&data), 50);
+        assert_eq!(p99(&data), 50);
+    }
+
+    #[test]
+    fn test_inode_for_existing_file() {
+        let tmp = tempfile::NamedTempFile::new().unwrap();
+        let inode = MetricsCollector::get_inode(tmp.path()).unwrap();
+        assert!(inode > 0, "inode should be non-zero");
+    }
+
+    #[test]
+    fn test_detect_rotation_no_rotation() {
+        let tmp = tempfile::NamedTempFile::new().unwrap();
+        let inode = MetricsCollector::get_inode(tmp.path()).unwrap();
+        assert!(!MetricsCollector::detect_rotation(inode, tmp.path()));
+    }
+
+    #[test]
+    fn test_detect_rotation_file_removed() {
+        let inode: u64 = 99999;
+        let result = MetricsCollector::detect_rotation(inode, Path::new("/no/such/file"));
+        assert!(result, "missing file should indicate rotation");
+    }
+
+    #[test]
+    fn test_clear_file_cache() {
+        let tmp = tempfile::NamedTempFile::new().unwrap();
+        let result = MetricsCollector::clear_file_cache(tmp.path());
+        assert!(
+            result.is_ok(),
+            "clear_file_cache should succeed on temp file: {result:?}"
+        );
+    }
+}

crates/bench/src/mmap_reader.rs

@@ -0,0 +1,752 @@
+// ─── mmap_reader.rs ──────────────────────────────────────────────────────────
+// mmap-based FileReaderBackend implementations with 5 variants for benchmarking.
+// Includes SIGBUS handler for file-truncation resilience and remap support
+// for growing-file scenarios.
+// ──────────────────────────────────────────────────────────────────────────────
+
+use std::fs::File;
+use std::io::BufReader;
+use std::path::Path;
+use std::sync::Once;
+use std::sync::atomic::{AtomicBool, AtomicPtr, AtomicU8, Ordering};
+
+use memmap2::{Advice, Mmap, MmapOptions, RemapOptions};
+use nix::sys::signal::{SaFlags, SigAction, SigHandler, SigSet, Signal, sigaction};
+
+use crate::FileReaderBackend;
+use crate::line_index::LineIndex;
+
+// ─── SIGBUS Handler ──────────────────────────────────────────────────────────
+//
+// Signal-safety architecture:
+// - Old handler state is stored in raw atomics (AtomicU8 + AtomicPtr) that are
+//   async-signal-safe to read — no OnceLock, Mutex, or other non-trivial abstractions
+//   in the signal handler path.
+// - Installation uses `Once` for idempotency and follows a strict publish-then-install
+//   sequence to close the "handler-active-but-state-unpublished" race window.
+
+/// Global flag set by the SIGBUS handler when a bus error is intercepted.
+/// Process-global: concurrent benchmarks that reset/check this flag may interfere.
+/// Currently acceptable because only `rotation` suite (sequential) uses it.
+static SIGBUS_OCCURRED: AtomicBool = AtomicBool::new(false);
+
+/// Discriminant values for old SIGBUS handler type.
+const HANDLER_NONE: u8 = 0;
+const HANDLER_DEFAULT: u8 = 1;
+const HANDLER_IGNORE: u8 = 2;
+const HANDLER_PLAIN: u8 = 3; // extern "C" fn(c_int)
+#[allow(
+    clippy::unseparated_literal_suffix,
+    reason = "clarity: this is the SA_SIGACTION variant"
+)]
+const HANDLER_SIGACTION: u8 = 4; // extern "C" fn(c_int, *mut siginfo_t, *mut c_void)
+
+/// Old SIGBUS handler type — raw atomic, async-signal-safe to read.
+static OLD_HANDLER_KIND: AtomicU8 = AtomicU8::new(HANDLER_NONE);
+
+/// Old SIGBUS handler function pointer — raw atomic, async-signal-safe to read.
+static OLD_HANDLER_PTR: AtomicPtr<std::ffi::c_void> = AtomicPtr::new(std::ptr::null_mut());
+
+/// Ensures `install_sigbus_handler` runs exactly once across all threads.
+static INSTALL_ONCE: Once = Once::new();
+
+/// Returns `true` if a SIGBUS was intercepted since the last reset.
+pub fn sigbus_flag() -> bool {
+    SIGBUS_OCCURRED.load(Ordering::SeqCst)
+}
+
+/// Resets the SIGBUS flag. Call before operations where you want to detect
+/// a fresh SIGBUS.
+pub fn reset_sigbus_flag() {
+    SIGBUS_OCCURRED.store(false, Ordering::SeqCst)
+}
+
+/// SIGBUS signal handler.
+///
+/// # Safety Constraints (signal handler context)
+/// - NO TLS access
+/// - NO memory allocation
+/// - NO lock acquisition
+/// - Only: AtomicBool store → mmap → raw atomic loads → chain
+extern "C" fn sigbus_handler(
+    sig: libc::c_int,
+    info: *mut libc::siginfo_t,
+    ctx: *mut std::ffi::c_void,
+) {
+    SIGBUS_OCCURRED.store(true, Ordering::SeqCst);
+
+    // Align fault address down to page boundary.
+    // si_addr is NOT guaranteed to be page-aligned; without this, mmap(MAP_FIXED)
+    // could replace the wrong page.
+    let addr = unsafe { (*info).si_addr() };
+    let aligned = (addr as usize & !0xFFF) as *mut libc::c_void;
+
+    // Map anonymous zero page at fault address to prevent crash on file truncation.
+    let result = unsafe {
+        libc::mmap(
+            aligned,
+            4096,
+            libc::PROT_READ,
+            libc::MAP_PRIVATE | libc::MAP_ANONYMOUS | libc::MAP_FIXED,
+            -1,
+            0,
+        )
+    };
+    if result == libc::MAP_FAILED {
+        // Chain to old handler via raw atomics (async-signal-safe).
+        let kind = OLD_HANDLER_KIND.load(Ordering::Acquire);
+        match kind {
+            HANDLER_PLAIN => {
+                let ptr = OLD_HANDLER_PTR.load(Ordering::Acquire);
+                if !ptr.is_null() {
+                    // SAFETY: ptr was derived from a valid function pointer
+                    // published by install_sigbus_handler before this handler was installed.
+                    let f: extern "C" fn(libc::c_int) = unsafe { std::mem::transmute(ptr) };
+                    f(sig);
+                } else {
+                    unsafe { libc::_exit(128 + sig) };
+                }
+            }
+            HANDLER_SIGACTION => {
+                let ptr = OLD_HANDLER_PTR.load(Ordering::Acquire);
+                if !ptr.is_null() {
+                    let f: extern "C" fn(libc::c_int, *mut libc::siginfo_t, *mut std::ffi::c_void) =
+                        unsafe { std::mem::transmute(ptr) };
+                    f(sig, info, ctx);
+                } else {
+                    unsafe { libc::_exit(128 + sig) };
+                }
+            }
+            _ => {
+                // HANDLER_NONE / HANDLER_DEFAULT / HANDLER_IGNORE — no safe chaining.
+                unsafe { libc::_exit(128 + sig) };
+            }
+        }
+    }
+}
+
+/// Install the SIGBUS handler via sigaction(). Chains to any previous handler.
+/// Uses `Once` to guarantee exactly-once installation — safe to call from
+/// multiple threads concurrently.
+///
+/// The install sequence closes the "install-then-publish" race by:
+/// 1. Querying the current handler via raw `libc::sigaction` (no modification).
+/// 2. Publishing old handler state to raw atomics (`Release` stores).
+/// 3. Installing our handler via `nix::sigaction`.
+///
+/// This ensures the atomics are readable *before* our handler can fire.
+fn install_sigbus_handler() {
+    INSTALL_ONCE.call_once(|| {
+        // Step 1: Query current SIGBUS disposition without changing it.
+        let mut old_act: libc::sigaction = unsafe { std::mem::zeroed() };
+        let ret = unsafe { libc::sigaction(libc::SIGBUS, std::ptr::null(), &mut old_act) };
+        if ret != 0 {
+            // Failed to query — skip installation entirely.
+            return;
+        }
+
+        // Step 2: Publish old handler to atomics BEFORE installing ours.
+        let is_siginfo = (old_act.sa_flags & libc::SA_SIGINFO) != 0;
+
+        let raw_usize = old_act.sa_sigaction as usize;
+
+        if raw_usize == 0 {
+            // SIG_DFL (null function pointer → default disposition)
+            OLD_HANDLER_KIND.store(HANDLER_DEFAULT, Ordering::Release);
+        } else if raw_usize == 1 {
+            // SIG_IGN (sentinel value 1 → ignore disposition)
+            OLD_HANDLER_KIND.store(HANDLER_IGNORE, Ordering::Release);
+        } else {
+            // Real handler function pointer
+            OLD_HANDLER_PTR.store(raw_usize as *mut std::ffi::c_void, Ordering::Release);
+            if is_siginfo {
+                OLD_HANDLER_KIND.store(HANDLER_SIGACTION, Ordering::Release);
+            } else {
+                OLD_HANDLER_KIND.store(HANDLER_PLAIN, Ordering::Release);
+            }
+        }
+
+        // Step 3: Install our handler. Atomics are already published,
+        // so the handler can safely read them from the moment it's installed.
+        let new_action = SigAction::new(
+            SigHandler::SigAction(sigbus_handler),
+            SaFlags::SA_SIGINFO,
+            SigSet::empty(),
+        );
+        let _ = unsafe { sigaction(Signal::SIGBUS, &new_action) };
+    });
+}
+
+// ─── Core MmapReader ─────────────────────────────────────────────────────────
+
+/// Core mmap-based reader. Holds the memory mapping, file handle, and
+/// line index. Used as the engine inside each variant wrapper.
+pub struct MmapReader {
+    mmap: Mmap,
+    #[allow(dead_code)]
+    file: File,
+    line_index: LineIndex,
+    file_size: u64,
+}
+
+impl MmapReader {
+    /// Open a file and create an mmap. Does NOT apply madvise.
+    /// The caller is responsible for applying the desired advice.
+    fn open_raw(path: &Path) -> std::io::Result<Self> {
+        let file = File::open(path)?;
+        let file_size = file.metadata()?.len();
+
+        let mmap = unsafe { Mmap::map(&file)? };
+
+        let line_index = {
+            let mut reader = BufReader::new(&file);
+            LineIndex::from_reader(&mut reader)?
+        };
+
+        Ok(Self {
+            mmap,
+            file,
+            line_index,
+            file_size,
+        })
+    }
+
+    /// Open with MmapOptions (for populate, etc.).
+    fn open_with_options(path: &Path, opts: &MmapOptions) -> std::io::Result<Self> {
+        let file = File::open(path)?;
+        let file_size = file.metadata()?.len();
+
+        let mmap = if file_size == 0 {
+            unsafe { Mmap::map(&file)? }
+        } else {
+            unsafe { opts.map(&file)? }
+        };
+
+        let line_index = {
+            let mut reader = BufReader::new(&file);
+            LineIndex::from_reader(&mut reader)?
+        };
+
+        Ok(Self {
+            mmap,
+            file,
+            line_index,
+            file_size,
+        })
+    }
+
+    /// Apply madvise to the entire mapping.
+    fn advise(&self, advice: Advice) {
+        let _ = self.mmap.advise(advice);
+    }
+
+    /// Grow the mmap to `new_size` bytes.
+    ///
+    /// # Safety
+    /// Caller must ensure no `&[u8]` references to the old mmap data exist.
+    /// After remap, any pointers derived from the old mapping may be invalid
+    /// (if the kernel moved the mapping).
+    pub unsafe fn remap(&mut self, new_size: usize) -> std::io::Result<()> {
+        unsafe {
+            self.mmap
+                .remap(new_size, RemapOptions::new().may_move(true))?;
+        }
+        self.file_size = new_size as u64;
+        Ok(())
+    }
+
+    #[inline]
+    pub fn file_size(&self) -> u64 {
+        self.file_size
+    }
+
+    #[inline]
+    pub fn total_lines(&self) -> usize {
+        self.line_index.line_count()
+    }
+
+    #[inline]
+    pub fn get_line(&self, idx: usize) -> Option<String> {
+        self.line_index
+            .get_line(&self.mmap, idx)
+            .map(|s| s.to_owned())
+    }
+
+    #[inline]
+    pub fn read_range(&self, offset: u64, len: usize) -> Option<Vec<u8>> {
+        let start = offset as usize;
+        let end = start.checked_add(len)?;
+        if end > self.mmap.len() {
+            return None;
+        }
+        Some(self.mmap[start..end].to_vec())
+    }
+}
+
+// ─── 5 Variants ──────────────────────────────────────────────────────────────
+
+/// Variant 1: Plain mmap, no madvise.
+pub struct MmapReaderPlain {
+    inner: MmapReader,
+}
+
+impl MmapReaderPlain {
+    pub fn remap(&mut self, new_size: usize) -> std::io::Result<()> {
+        unsafe { self.inner.remap(new_size) }
+    }
+}
+
+impl FileReaderBackend for MmapReaderPlain {
+    fn name(&self) -> &str {
+        "mmap_plain"
+    }
+
+    fn open(path: &Path) -> std::io::Result<Self>
+    where
+        Self: Sized,
+    {
+        install_sigbus_handler();
+        let inner = MmapReader::open_raw(path)?;
+        Ok(Self { inner })
+    }
+
+    fn file_size(&self) -> u64 {
+        self.inner.file_size()
+    }
+
+    fn total_lines(&self) -> usize {
+        self.inner.total_lines()
+    }
+
+    fn get_line(&self, idx: usize) -> Option<String> {
+        self.inner.get_line(idx)
+    }
+
+    fn read_range(&self, offset: u64, len: usize) -> Option<Vec<u8>> {
+        self.inner.read_range(offset, len)
+    }
+
+    fn close(self) {}
+}
+
+/// Variant 2: mmap with MADV_SEQUENTIAL — optimal for sequential scan (index build).
+pub struct MmapReaderSequential {
+    inner: MmapReader,
+}
+
+impl FileReaderBackend for MmapReaderSequential {
+    fn name(&self) -> &str {
+        "mmap_sequential"
+    }
+
+    fn open(path: &Path) -> std::io::Result<Self>
+    where
+        Self: Sized,
+    {
+        install_sigbus_handler();
+        let inner = MmapReader::open_raw(path)?;
+        inner.advise(Advice::Sequential);
+        Ok(Self { inner })
+    }
+
+    fn file_size(&self) -> u64 {
+        self.inner.file_size()
+    }
+
+    fn total_lines(&self) -> usize {
+        self.inner.total_lines()
+    }
+
+    fn get_line(&self, idx: usize) -> Option<String> {
+        self.inner.get_line(idx)
+    }
+
+    fn read_range(&self, offset: u64, len: usize) -> Option<Vec<u8>> {
+        self.inner.read_range(offset, len)
+    }
+
+    fn close(self) {}
+}
+
+/// Variant 3: mmap with MADV_RANDOM — optimal for random line access after index is built.
+pub struct MmapReaderRandom {
+    inner: MmapReader,
+}
+
+impl FileReaderBackend for MmapReaderRandom {
+    fn name(&self) -> &str {
+        "mmap_random"
+    }
+
+    fn open(path: &Path) -> std::io::Result<Self>
+    where
+        Self: Sized,
+    {
+        install_sigbus_handler();
+        let inner = MmapReader::open_raw(path)?;
+        inner.advise(Advice::Random);
+        Ok(Self { inner })
+    }
+
+    fn file_size(&self) -> u64 {
+        self.inner.file_size()
+    }
+
+    fn total_lines(&self) -> usize {
+        self.inner.total_lines()
+    }
+
+    fn get_line(&self, idx: usize) -> Option<String> {
+        self.inner.get_line(idx)
+    }
+
+    fn read_range(&self, offset: u64, len: usize) -> Option<Vec<u8>> {
+        self.inner.read_range(offset, len)
+    }
+
+    fn close(self) {}
+}
+
+/// Variant 4: mmap with MAP_POPULATE — pre-fault all pages at mmap time.
+/// Trades higher upfront cost for smoother subsequent access.
+pub struct MmapReaderPopulate {
+    inner: MmapReader,
+}
+
+impl FileReaderBackend for MmapReaderPopulate {
+    fn name(&self) -> &str {
+        "mmap_populate"
+    }
+
+    fn open(path: &Path) -> std::io::Result<Self>
+    where
+        Self: Sized,
+    {
+        install_sigbus_handler();
+        let mut opts = MmapOptions::new();
+        opts.populate();
+        let inner = MmapReader::open_with_options(path, &opts)?;
+        Ok(Self { inner })
+    }
+
+    fn file_size(&self) -> u64 {
+        self.inner.file_size()
+    }
+
+    fn total_lines(&self) -> usize {
+        self.inner.total_lines()
+    }
+
+    fn get_line(&self, idx: usize) -> Option<String> {
+        self.inner.get_line(idx)
+    }
+
+    fn read_range(&self, offset: u64, len: usize) -> Option<Vec<u8>> {
+        self.inner.read_range(offset, len)
+    }
+
+    fn close(self) {}
+}
+
+/// Variant 5: Phase-aware — MADV_SEQUENTIAL during index build, then MADV_RANDOM
+/// for line access. Best of both worlds for the read-index-then-query pattern.
+pub struct MmapReaderPhaseAware {
+    inner: MmapReader,
+}
+
+impl FileReaderBackend for MmapReaderPhaseAware {
+    fn name(&self) -> &str {
+        "mmap_phase_aware"
+    }
+
+    fn open(path: &Path) -> std::io::Result<Self>
+    where
+        Self: Sized,
+    {
+        install_sigbus_handler();
+        let inner = MmapReader::open_raw(path)?;
+        // Index build used sequential streaming via BufReader. Now switch to random.
+        inner.advise(Advice::Random);
+        Ok(Self { inner })
+    }
+
+    fn file_size(&self) -> u64 {
+        self.inner.file_size()
+    }
+
+    fn total_lines(&self) -> usize {
+        self.inner.total_lines()
+    }
+
+    fn get_line(&self, idx: usize) -> Option<String> {
+        self.inner.get_line(idx)
+    }
+
+    fn read_range(&self, offset: u64, len: usize) -> Option<Vec<u8>> {
+        self.inner.read_range(offset, len)
+    }
+
+    fn close(self) {}
+}
+
+// ─── Tests ───────────────────────────────────────────────────────────────────
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::io::Write as _;
+    use tempfile::NamedTempFile;
+
+    fn create_temp_file(content: &[u8]) -> NamedTempFile {
+        let mut f = NamedTempFile::new().unwrap();
+        f.write_all(content).unwrap();
+        f.flush().unwrap();
+        f
+    }
+
+    #[test]
+    fn test_plain_open_and_read_lines() {
+        let f = create_temp_file(b"hello\nworld\nfoo\n");
+        let reader = MmapReaderPlain::open(f.path()).unwrap();
+        assert_eq!(reader.name(), "mmap_plain");
+        assert_eq!(reader.total_lines(), 3);
+        assert_eq!(reader.get_line(0), Some("hello".to_owned()));
+        assert_eq!(reader.get_line(1), Some("world".to_owned()));
+        assert_eq!(reader.get_line(2), Some("foo".to_owned()));
+        assert_eq!(reader.get_line(3), None);
+        reader.close();
+    }
+
+    #[test]
+    fn test_plain_out_of_bounds_returns_none() {
+        let f = create_temp_file(b"line1\nline2\n");
+        let reader = MmapReaderPlain::open(f.path()).unwrap();
+        assert_eq!(reader.total_lines(), 2);
+        assert_eq!(reader.get_line(100), None);
+        assert_eq!(reader.get_line(usize::MAX), None);
+        reader.close();
+    }
+
+    #[test]
+    fn test_plain_read_range() {
+        let content = b"hello world\nfoo bar\n";
+        let f = create_temp_file(content);
+        let reader = MmapReaderPlain::open(f.path()).unwrap();
+        let range = reader.read_range(0, 5).unwrap();
+        assert_eq!(&range, b"hello");
+        let range = reader.read_range(6, 5).unwrap();
+        assert_eq!(&range, b"world");
+        assert_eq!(reader.read_range(0, 1000), None);
+        reader.close();
+    }
+
+    #[test]
+    fn test_plain_empty_file() {
+        let f = create_temp_file(b"");
+        let reader = MmapReaderPlain::open(f.path()).unwrap();
+        assert_eq!(reader.total_lines(), 0);
+        assert_eq!(reader.file_size(), 0);
+        assert_eq!(reader.get_line(0), None);
+        reader.close();
+    }
+
+    #[test]
+    fn test_sequential_variant() {
+        let f = create_temp_file(b"alpha\nbeta\ngamma\n");
+        let reader = MmapReaderSequential::open(f.path()).unwrap();
+        assert_eq!(reader.name(), "mmap_sequential");
+        assert_eq!(reader.total_lines(), 3);
+        assert_eq!(reader.get_line(1), Some("beta".to_owned()));
+        reader.close();
+    }
+
+    #[test]
+    fn test_random_variant() {
+        let f = create_temp_file(b"one\ntwo\nthree\n");
+        let reader = MmapReaderRandom::open(f.path()).unwrap();
+        assert_eq!(reader.name(), "mmap_random");
+        assert_eq!(reader.total_lines(), 3);
+        assert_eq!(reader.get_line(2), Some("three".to_owned()));
+        reader.close();
+    }
+
+    #[test]
+    fn test_populate_variant() {
+        let f = create_temp_file(b"x\ny\nz\n");
+        let reader = MmapReaderPopulate::open(f.path()).unwrap();
+        assert_eq!(reader.name(), "mmap_populate");
+        assert_eq!(reader.total_lines(), 3);
+        assert_eq!(reader.get_line(0), Some("x".to_owned()));
+        reader.close();
+    }
+
+    #[test]
+    fn test_phase_aware_variant() {
+        let f = create_temp_file(b"a\nb\nc\n");
+        let reader = MmapReaderPhaseAware::open(f.path()).unwrap();
+        assert_eq!(reader.name(), "mmap_phase_aware");
+        assert_eq!(reader.total_lines(), 3);
+        assert_eq!(reader.get_line(1), Some("b".to_owned()));
+        reader.close();
+    }
+
+    #[test]
+    fn test_file_size() {
+        let content = b"hello world";
+        let f = create_temp_file(content);
+        let reader = MmapReaderPlain::open(f.path()).unwrap();
+        assert_eq!(reader.file_size(), content.len() as u64);
+        reader.close();
+    }
+
+    #[test]
+    fn test_sigbus_flag_api() {
+        reset_sigbus_flag();
+        assert!(!sigbus_flag());
+    }
+
+    #[test]
+    fn test_concurrent_open_installs_handler_once() {
+        let f = create_temp_file(b"line1\nline2\nline3\n");
+        let path = f.path().to_owned();
+        let num_threads = 8;
+
+        let handles: Vec<_> = (0..num_threads)
+            .map(|thread_id| {
+                let path = path.clone();
+                std::thread::spawn(move || {
+                    let reader = match thread_id % 5 {
+                        0 => {
+                            let r = MmapReaderPlain::open(&path).unwrap();
+                            assert_eq!(r.total_lines(), 3);
+                            r.close();
+                            "plain"
+                        }
+                        1 => {
+                            let r = MmapReaderSequential::open(&path).unwrap();
+                            assert_eq!(r.total_lines(), 3);
+                            r.close();
+                            "sequential"
+                        }
+                        2 => {
+                            let r = MmapReaderRandom::open(&path).unwrap();
+                            assert_eq!(r.total_lines(), 3);
+                            r.close();
+                            "random"
+                        }
+                        3 => {
+                            let r = MmapReaderPopulate::open(&path).unwrap();
+                            assert_eq!(r.total_lines(), 3);
+                            r.close();
+                            "populate"
+                        }
+                        _ => {
+                            let r = MmapReaderPhaseAware::open(&path).unwrap();
+                            assert_eq!(r.total_lines(), 3);
+                            r.close();
+                            "phase_aware"
+                        }
+                    };
+                    reader.to_owned()
+                })
+            })
+            .collect();
+
+        let results: Vec<String> = handles
+            .into_iter()
+            .map(|h| h.join().expect("thread panicked"))
+            .collect();
+        assert_eq!(results.len(), num_threads);
+    }
+
+    #[test]
+    fn test_no_trailing_newline() {
+        let f = create_temp_file(b"line1\nline2");
+        let reader = MmapReaderPlain::open(f.path()).unwrap();
+        assert_eq!(reader.total_lines(), 2);
+        assert_eq!(reader.get_line(0), Some("line1".to_owned()));
+        assert_eq!(reader.get_line(1), Some("line2".to_owned()));
+        reader.close();
+    }
+
+    /// Diagnostic test: measure how long each stage of `open_raw()` takes
+    /// on a large file (e.g. the 5GB extreme.log).
+    /// Run with: cargo test -p log-viewer-bench --release -- --nocapture diag_open_stages
+    #[test]
+    fn diag_open_stages() {
+        let path = std::path::Path::new("/tmp/test-logviewer/extreme.log");
+        if !path.exists() {
+            eprintln!("SKIP: test file not found");
+            return;
+        }
+
+        // Stage 1: File::open + metadata
+        let t0 = std::time::Instant::now();
+        let file = File::open(path).unwrap();
+        let file_size = file.metadata().unwrap().len();
+        let stage1 = t0.elapsed();
+        eprintln!(
+            "Stage 1 - File::open + metadata: {:.2}ms ({:.1}GB)",
+            stage1.as_secs_f64() * 1000.0,
+            file_size as f64 / 1073741824.0
+        );
+
+        // Stage 2: mmap::map
+        let t1 = std::time::Instant::now();
+        let _mmap = unsafe { Mmap::map(&file) }.unwrap();
+        let stage2 = t1.elapsed();
+        eprintln!(
+            "Stage 2 - mmap::map:                {:.2}ms",
+            stage2.as_secs_f64() * 1000.0
+        );
+        drop(_mmap);
+
+        // Stage 3: LineIndex::from_reader via BufReader (default 8KB buffer)
+        let t2 = std::time::Instant::now();
+        let mut reader = BufReader::new(&file);
+        let line_index = crate::line_index::LineIndex::from_reader(&mut reader).unwrap();
+        let stage3 = t2.elapsed();
+        eprintln!(
+            "Stage 3 - LineIndex::from_reader:   {:.2}ms ({} lines, {} sampled_offsets)",
+            stage3.as_secs_f64() * 1000.0,
+            line_index.line_count(),
+            line_index.sampled_offsets.len()
+        );
+
+        // Stage 3b: Try with 1MB buffer
+        let file2 = File::open(path).unwrap();
+        let t2b = std::time::Instant::now();
+        let mut reader2 = BufReader::with_capacity(1024 * 1024, &file2);
+        let line_index2 = crate::line_index::LineIndex::from_reader(&mut reader2).unwrap();
+        let stage3b = t2b.elapsed();
+        eprintln!(
+            "Stage 3b - LineIndex (1MB buffer):   {:.2}ms",
+            stage3b.as_secs_f64() * 1000.0
+        );
+
+        eprintln!("\n=== 瓶颈分析 ===");
+        let total_ms = stage1.as_secs_f64() * 1000.0
+            + stage2.as_secs_f64() * 1000.0
+            + stage3.as_secs_f64() * 1000.0;
+        let total_dur = stage1 + stage2 + stage3;
+        eprintln!("Total ~{:.0}ms", total_ms);
+        eprintln!(
+            "  File::open:  {:.1}% ({:.0}ms)",
+            stage1.as_secs_f64() * 1000.0 / total_dur.as_secs_f64() / 1000.0 * 100.0,
+            stage1.as_secs_f64() * 1000.0
+        );
+        eprintln!(
+            "  mmap::map:   {:.1}% ({:.0}ms)",
+            stage2.as_secs_f64() * 1000.0 / total_dur.as_secs_f64() / 1000.0 * 100.0,
+            stage2.as_secs_f64() * 1000.0
+        );
+        eprintln!(
+            "  from_reader: {:.1}% ({:.0}ms)",
+            stage3.as_secs_f64() * 1000.0 / total_dur.as_secs_f64() / 1000.0 * 100.0,
+            stage3.as_secs_f64() * 1000.0
+        );
+
+        // Suppress unused warnings
+        let _ = line_index2;
+    }
+}

crates/bench/src/pread_reader.rs

@@ -0,0 +1,533 @@
+// ─── pread_reader.rs ──────────────────────────────────────────────────────────
+// pread-based FileReaderBackend implementations with 3 variants for benchmarking.
+// Uses ReadCache for 4KB block caching to reduce syscalls, and posix_fadvise
+// for kernel readahead control.
+//
+// CRITICAL: get_line() is custom — it does NOT use LineIndex::get_line() which
+// requires a full &[u8] data slice. Instead it uses sampled_offsets directly
+// and reads on-demand via pread.
+// ──────────────────────────────────────────────────────────────────────────────
+
+use std::cell::RefCell;
+use std::fs::File;
+use std::io::{BufReader, Seek as _, SeekFrom};
+use std::os::unix::fs::FileExt;
+use std::os::unix::io::AsRawFd;
+use std::path::Path;
+
+use crate::FileReaderBackend;
+use crate::line_index::LineIndex;
+
+const BLOCK_SIZE: usize = 256;
+const CACHE_CHUNK: usize = 4096;
+
+// ─── ReadCache ────────────────────────────────────────────────────────────────
+
+/// Single-block read cache. Reduces syscalls by caching the last read.
+/// Typical cache hit: sequential get_line() calls within the same 4KB block.
+struct ReadCache {
+    buf: Vec<u8>,
+    buf_offset: u64,
+    buf_len: usize,
+}
+
+impl ReadCache {
+    fn new() -> Self {
+        Self {
+            buf: vec![0u8; CACHE_CHUNK],
+            buf_offset: 0,
+            buf_len: 0,
+        }
+    }
+
+    /// Read `len` bytes starting at `offset`. Returns a slice into the cache.
+    /// On cache hit (range fully within cached block), no syscall needed.
+    /// On miss, performs a `read_exact_at` syscall.
+    fn invalidate(&mut self) {
+        self.buf_len = 0;
+    }
+
+    fn get(&mut self, file: &File, offset: u64, len: usize) -> std::io::Result<&[u8]> {
+        let end = offset + len as u64;
+        if offset >= self.buf_offset && end <= self.buf_offset + self.buf_len as u64 {
+            let start = (offset - self.buf_offset) as usize;
+            return Ok(&self.buf[start..start + len]);
+        }
+        let alloc_len = CACHE_CHUNK.max(len);
+        self.buf.resize(alloc_len, 0);
+        file.read_exact_at(&mut self.buf[..len], offset)?;
+        self.buf_offset = offset;
+        self.buf_len = len;
+        Ok(&self.buf[..len])
+    }
+}
+
+// ─── PreadReaderCore ──────────────────────────────────────────────────────────
+
+/// Core pread-based reader. Uses ReadCache and sparse LineIndex for on-demand
+/// line retrieval without mmap.
+pub struct PreadReaderCore {
+    file: File,
+    line_index: LineIndex,
+    file_size: u64,
+    cache: RefCell<ReadCache>,
+}
+
+impl PreadReaderCore {
+    /// Open file, build line index via streaming BufReader, prepare pread reader.
+    fn open_raw(path: &Path) -> std::io::Result<Self> {
+        let file = File::open(path)?;
+        let file_size = file.metadata()?.len();
+
+        let line_index = {
+            let mut reader = BufReader::new(&file);
+            LineIndex::from_reader(&mut reader)?
+        };
+
+        Ok(Self {
+            file,
+            line_index,
+            file_size,
+            cache: RefCell::new(ReadCache::new()),
+        })
+    }
+
+    /// Apply posix_fadvise to the entire file.
+    fn advise(&self, advice: libc::c_int) -> std::io::Result<()> {
+        let fd = self.file.as_raw_fd();
+        let ret = unsafe { libc::posix_fadvise(fd, 0, self.file_size as i64, advice) };
+        if ret != 0 {
+            return Err(std::io::Error::from_raw_os_error(ret));
+        }
+        Ok(())
+    }
+
+    /// Custom get_line using sparse index + pread.
+    /// Does NOT use LineIndex::get_line() (which needs full &[u8] slice).
+    ///
+    /// Algorithm:
+    /// 1. Look up sampled_offsets[block] to get approximate byte position
+    /// 2. Scan forward through offset_in_block newlines via memchr
+    /// 3. Collect bytes from that position until next newline (may span 4KB blocks)
+    fn get_line_impl(&self, idx: usize) -> Option<String> {
+        let total = self.line_index.total_lines as usize;
+        if idx >= total || total == 0 {
+            return None;
+        }
+
+        let block = idx / BLOCK_SIZE;
+        let offset_in_block = idx % BLOCK_SIZE;
+
+        let mut cache = self.cache.borrow_mut();
+
+        // Phase 1: Find byte position of the start of line `idx`.
+        // Start at sampled_offsets[block], scan forward through `offset_in_block` newlines.
+        let start_offset = self.line_index.sampled_offsets[block];
+        let mut pos = start_offset;
+        let mut newlines_found = 0;
+
+        while newlines_found < offset_in_block {
+            let remaining = self.file_size.saturating_sub(pos) as usize;
+            if remaining == 0 {
+                return None;
+            }
+            let to_read = CACHE_CHUNK.min(remaining);
+            let data = cache.get(&self.file, pos, to_read).ok()?;
+
+            for byte_pos in memchr::memchr_iter(b'\n', data) {
+                newlines_found += 1;
+                if newlines_found == offset_in_block {
+                    pos = pos + byte_pos as u64 + 1;
+                    break;
+                }
+            }
+
+            if newlines_found < offset_in_block {
+                pos += to_read as u64;
+            }
+        }
+
+        // Phase 2: Collect bytes from `pos` until next newline or EOF.
+        // Line data may span multiple 4KB cache blocks.
+        let mut result = Vec::new();
+
+        while pos < self.file_size {
+            let remaining = (self.file_size - pos) as usize;
+            let to_read = CACHE_CHUNK.min(remaining);
+            let data = cache.get(&self.file, pos, to_read).ok()?;
+
+            match memchr::memchr(b'\n', data) {
+                Some(rel) => {
+                    result.extend_from_slice(&data[..rel]);
+                    break;
+                }
+                None => {
+                    result.extend_from_slice(data);
+                    pos += to_read as u64;
+                }
+            }
+        }
+
+        String::from_utf8(result)
+            .ok()
+            .map(|s| s.trim_end_matches(['\r', '\n']).to_owned())
+    }
+
+    /// Read a raw byte range from the file using pread.
+    fn read_range_impl(&self, offset: u64, len: usize) -> Option<Vec<u8>> {
+        let end = offset.checked_add(len as u64)?;
+        if end > self.file_size {
+            return None;
+        }
+        let mut buf = vec![0u8; len];
+        self.file.read_exact_at(&mut buf, offset).ok()?;
+        Some(buf)
+    }
+
+    #[inline]
+    pub fn file_size(&self) -> u64 {
+        self.file_size
+    }
+
+    #[inline]
+    pub fn total_lines(&self) -> usize {
+        self.line_index.line_count()
+    }
+
+    pub fn refresh_index(&mut self) -> std::io::Result<()> {
+        let new_size = self.file.metadata()?.len();
+        self.file.seek(SeekFrom::Start(0))?;
+        let new_index = {
+            let mut reader = BufReader::new(&self.file);
+            LineIndex::from_reader(&mut reader)?
+        };
+        self.file_size = new_size;
+        self.line_index = new_index;
+        self.cache.get_mut().invalidate();
+        Ok(())
+    }
+}
+
+// ─── 3 Variants ───────────────────────────────────────────────────────────────
+
+/// Variant 1: Plain pread, no fadvise.
+pub struct PreadReaderPlain {
+    inner: PreadReaderCore,
+}
+
+impl PreadReaderPlain {
+    pub fn refresh_index(&mut self) -> std::io::Result<()> {
+        self.inner.refresh_index()
+    }
+}
+
+impl FileReaderBackend for PreadReaderPlain {
+    fn name(&self) -> &str {
+        "pread_plain"
+    }
+
+    fn open(path: &Path) -> std::io::Result<Self>
+    where
+        Self: Sized,
+    {
+        let inner = PreadReaderCore::open_raw(path)?;
+        Ok(Self { inner })
+    }
+
+    fn file_size(&self) -> u64 {
+        self.inner.file_size()
+    }
+
+    fn total_lines(&self) -> usize {
+        self.inner.total_lines()
+    }
+
+    fn get_line(&self, idx: usize) -> Option<String> {
+        self.inner.get_line_impl(idx)
+    }
+
+    fn read_range(&self, offset: u64, len: usize) -> Option<Vec<u8>> {
+        self.inner.read_range_impl(offset, len)
+    }
+
+    fn close(self) {}
+}
+
+/// Variant 2: pread with POSIX_FADV_RANDOM — disable kernel readahead.
+/// Best for random-access line lookup patterns.
+pub struct PreadReaderRandom {
+    inner: PreadReaderCore,
+}
+
+impl FileReaderBackend for PreadReaderRandom {
+    fn name(&self) -> &str {
+        "pread_random"
+    }
+
+    fn open(path: &Path) -> std::io::Result<Self>
+    where
+        Self: Sized,
+    {
+        let inner = PreadReaderCore::open_raw(path)?;
+        inner.advise(libc::POSIX_FADV_RANDOM)?;
+        Ok(Self { inner })
+    }
+
+    fn file_size(&self) -> u64 {
+        self.inner.file_size()
+    }
+
+    fn total_lines(&self) -> usize {
+        self.inner.total_lines()
+    }
+
+    fn get_line(&self, idx: usize) -> Option<String> {
+        self.inner.get_line_impl(idx)
+    }
+
+    fn read_range(&self, offset: u64, len: usize) -> Option<Vec<u8>> {
+        self.inner.read_range_impl(offset, len)
+    }
+
+    fn close(self) {}
+}
+
+/// Variant 3: pread with POSIX_FADV_SEQUENTIAL — aggressive kernel readahead.
+/// Best for sequential scan patterns.
+pub struct PreadReaderSequential {
+    inner: PreadReaderCore,
+}
+
+impl FileReaderBackend for PreadReaderSequential {
+    fn name(&self) -> &str {
+        "pread_sequential"
+    }
+
+    fn open(path: &Path) -> std::io::Result<Self>
+    where
+        Self: Sized,
+    {
+        let inner = PreadReaderCore::open_raw(path)?;
+        inner.advise(libc::POSIX_FADV_SEQUENTIAL)?;
+        Ok(Self { inner })
+    }
+
+    fn file_size(&self) -> u64 {
+        self.inner.file_size()
+    }
+
+    fn total_lines(&self) -> usize {
+        self.inner.total_lines()
+    }
+
+    fn get_line(&self, idx: usize) -> Option<String> {
+        self.inner.get_line_impl(idx)
+    }
+
+    fn read_range(&self, offset: u64, len: usize) -> Option<Vec<u8>> {
+        self.inner.read_range_impl(offset, len)
+    }
+
+    fn close(self) {}
+}
+
+// ─── Tests ────────────────────────────────────────────────────────────────────
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::io::Write as _;
+    use tempfile::NamedTempFile;
+
+    fn create_temp_file(content: &[u8]) -> NamedTempFile {
+        let mut f = NamedTempFile::new().unwrap();
+        f.write_all(content).unwrap();
+        f.flush().unwrap();
+        f
+    }
+
+    #[test]
+    fn test_plain_open_and_read_lines() {
+        let f = create_temp_file(b"hello\nworld\nfoo\n");
+        let reader = PreadReaderPlain::open(f.path()).unwrap();
+        assert_eq!(reader.name(), "pread_plain");
+        assert_eq!(reader.total_lines(), 3);
+        assert_eq!(reader.get_line(0), Some("hello".to_owned()));
+        assert_eq!(reader.get_line(1), Some("world".to_owned()));
+        assert_eq!(reader.get_line(2), Some("foo".to_owned()));
+        assert_eq!(reader.get_line(3), None);
+        reader.close();
+    }
+
+    #[test]
+    fn test_plain_out_of_bounds_returns_none() {
+        let f = create_temp_file(b"line1\nline2\n");
+        let reader = PreadReaderPlain::open(f.path()).unwrap();
+        assert_eq!(reader.total_lines(), 2);
+        assert_eq!(reader.get_line(100), None);
+        assert_eq!(reader.get_line(usize::MAX), None);
+        reader.close();
+    }
+
+    #[test]
+    fn test_plain_read_range() {
+        let content = b"hello world\nfoo bar\n";
+        let f = create_temp_file(content);
+        let reader = PreadReaderPlain::open(f.path()).unwrap();
+        let range = reader.read_range(0, 5).unwrap();
+        assert_eq!(&range, b"hello");
+        let range = reader.read_range(6, 5).unwrap();
+        assert_eq!(&range, b"world");
+        assert_eq!(reader.read_range(0, 1000), None);
+        reader.close();
+    }
+
+    #[test]
+    fn test_plain_empty_file() {
+        let f = create_temp_file(b"");
+        let reader = PreadReaderPlain::open(f.path()).unwrap();
+        assert_eq!(reader.total_lines(), 0);
+        assert_eq!(reader.file_size(), 0);
+        assert_eq!(reader.get_line(0), None);
+        reader.close();
+    }
+
+    #[test]
+    fn test_random_variant() {
+        let f = create_temp_file(b"one\ntwo\nthree\n");
+        let reader = PreadReaderRandom::open(f.path()).unwrap();
+        assert_eq!(reader.name(), "pread_random");
+        assert_eq!(reader.total_lines(), 3);
+        assert_eq!(reader.get_line(2), Some("three".to_owned()));
+        reader.close();
+    }
+
+    #[test]
+    fn test_sequential_variant() {
+        let f = create_temp_file(b"alpha\nbeta\ngamma\n");
+        let reader = PreadReaderSequential::open(f.path()).unwrap();
+        assert_eq!(reader.name(), "pread_sequential");
+        assert_eq!(reader.total_lines(), 3);
+        assert_eq!(reader.get_line(1), Some("beta".to_owned()));
+        reader.close();
+    }
+
+    #[test]
+    fn test_file_size() {
+        let content = b"hello world";
+        let f = create_temp_file(content);
+        let reader = PreadReaderPlain::open(f.path()).unwrap();
+        assert_eq!(reader.file_size(), content.len() as u64);
+        reader.close();
+    }
+
+    #[test]
+    fn test_no_trailing_newline() {
+        let f = create_temp_file(b"line1\nline2");
+        let reader = PreadReaderPlain::open(f.path()).unwrap();
+        assert_eq!(reader.total_lines(), 2);
+        assert_eq!(reader.get_line(0), Some("line1".to_owned()));
+        assert_eq!(reader.get_line(1), Some("line2".to_owned()));
+        reader.close();
+    }
+
+    #[test]
+    fn test_line_spanning_cache_boundary() {
+        // Create a line that spans a 4KB boundary
+        let mut content = vec![b'a'; 4090];
+        content.push(b'\n');
+        content.extend_from_slice(b"target\n");
+        let f = create_temp_file(&content);
+        let reader = PreadReaderPlain::open(f.path()).unwrap();
+        assert_eq!(reader.get_line(0).unwrap().len(), 4090);
+        assert_eq!(reader.get_line(1), Some("target".to_owned()));
+        assert_eq!(reader.get_line(2), None);
+        reader.close();
+    }
+
+    #[test]
+    fn test_many_lines_random_access() {
+        // Create 300 lines to test block boundary (256 lines per block)
+        let mut content = String::new();
+        for i in 0..300 {
+            content.push_str(&format!("line_{}\n", i));
+        }
+        let f = create_temp_file(content.as_bytes());
+        let reader = PreadReaderPlain::open(f.path()).unwrap();
+        assert_eq!(reader.total_lines(), 300);
+        // Test lines across the 256-line block boundary
+        assert_eq!(reader.get_line(0), Some("line_0".to_owned()));
+        assert_eq!(reader.get_line(255), Some("line_255".to_owned()));
+        assert_eq!(reader.get_line(256), Some("line_256".to_owned()));
+        assert_eq!(reader.get_line(299), Some("line_299".to_owned()));
+        assert_eq!(reader.get_line(300), None);
+        reader.close();
+    }
+
+    #[test]
+    fn test_refresh_index_sees_appended_lines() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("refresh_test.log");
+
+        // Phase 1: write 3 initial lines
+        {
+            let mut f = std::fs::File::create(&path).unwrap();
+            std::io::Write::write_all(&mut f, b"alpha\nbeta\ngamma\n").unwrap();
+        }
+
+        let mut reader = PreadReaderPlain::open(&path).unwrap();
+        assert_eq!(reader.total_lines(), 3);
+        assert_eq!(reader.get_line(0), Some("alpha".to_owned()));
+        assert_eq!(
+            reader.get_line(3),
+            None,
+            "should be out of bounds before append"
+        );
+
+        // Phase 2: append 2 more lines
+        {
+            use std::io::Write as _;
+            let mut f = std::fs::OpenOptions::new()
+                .append(true)
+                .open(&path)
+                .unwrap();
+            f.write_all(b"delta\nepsilon\n").unwrap();
+        }
+
+        // Still stale — refresh needed
+        assert_eq!(reader.get_line(3), None, "stale index before refresh");
+
+        reader.refresh_index().unwrap();
+        assert_eq!(reader.total_lines(), 5);
+        assert_eq!(reader.get_line(3), Some("delta".to_owned()));
+        assert_eq!(reader.get_line(4), Some("epsilon".to_owned()));
+        assert_eq!(reader.get_line(5), None);
+
+        reader.close();
+    }
+
+    #[test]
+    fn test_refresh_index_no_change_is_noop() {
+        let f = create_temp_file(b"one\ntwo\n");
+        let mut reader = PreadReaderPlain::open(f.path()).unwrap();
+        assert_eq!(reader.total_lines(), 2);
+
+        reader.refresh_index().unwrap();
+        assert_eq!(reader.total_lines(), 2);
+        assert_eq!(reader.get_line(0), Some("one".to_owned()));
+        assert_eq!(reader.get_line(1), Some("two".to_owned()));
+
+        reader.close();
+    }
+
+    #[test]
+    fn test_truncation_graceful_error() {
+        // Verify that truncated file access returns None/Err instead of panicking
+        let f = create_temp_file(b"hello\nworld\n");
+        let reader = PreadReaderPlain::open(f.path()).unwrap();
+        assert_eq!(reader.get_line(0), Some("hello".to_owned()));
+        // Truncate behind the reader — pread may return zeros or error
+        let _ = reader.get_line(1); // must not panic
+        reader.close();
+    }
+}

crates/bench/src/report.rs

@@ -0,0 +1,301 @@
+use crate::metrics;
+use crate::types::BenchmarkResult;
+
+fn format_rss_mb(kb: u64) -> String {
+    format!("{:.1}MB", kb as f64 / 1024.0)
+}
+
+fn format_faults(count: u64) -> String {
+    if count < 1000 {
+        count.to_string()
+    } else {
+        let s = count.to_string();
+        let mut result = String::new();
+        for (i, c) in s.chars().rev().enumerate() {
+            if i > 0 && i % 3 == 0 {
+                result.push(',');
+            }
+            result.push(c);
+        }
+        result.chars().rev().collect()
+    }
+}
+
+// Format benchmark results as Markdown, grouped by category.
+pub fn format_report(results: &[BenchmarkResult]) -> String {
+    let mut report = String::new();
+    report.push_str("# Benchmark: mmap vs pread\n\n");
+
+    let mut categories: std::collections::BTreeMap<&str, Vec<&BenchmarkResult>> =
+        std::collections::BTreeMap::new();
+    for r in results {
+        categories.entry(&r.category).or_default().push(r);
+    }
+
+    for (category, category_results) in &categories {
+        report.push_str(&format!("## {}\n\n", capitalize(category)));
+
+        let mut tests: std::collections::BTreeMap<&str, Vec<&BenchmarkResult>> =
+            std::collections::BTreeMap::new();
+        for r in category_results {
+            tests.entry(&r.test_name).or_default().push(r);
+        }
+
+        let mut variants: Vec<(String, String)> = Vec::new();
+        for r in category_results {
+            if !variants
+                .iter()
+                .any(|(b, v)| b == &r.backend && v == &r.variant)
+            {
+                variants.push((r.backend.clone(), r.variant.clone()));
+            }
+        }
+        variants.sort();
+
+        report.push_str("### Latency\n\n");
+        report.push_str("| Test |");
+        for (backend, variant) in &variants {
+            report.push_str(&format!(" {} ({}) |", backend, variant));
+        }
+        report.push_str(" Winner |\n");
+
+        report.push_str("|------|");
+        for _ in &variants {
+            report.push_str("------|");
+        }
+        report.push_str("--------|\n");
+
+        for (test_name, test_results) in &tests {
+            report.push_str(&format!("| {} |", test_name));
+
+            let mut all_means: Vec<f64> = Vec::new();
+            for (backend, variant) in &variants {
+                if let Some(r) = test_results
+                    .iter()
+                    .find(|r| r.backend == *backend && r.variant == *variant)
+                {
+                    let avg = if r.latency_us.is_empty() {
+                        0.0
+                    } else {
+                        metrics::mean(&r.latency_us)
+                    };
+                    all_means.push(avg);
+                }
+            }
+            let use_ms = all_means.iter().any(|&v| v >= 1000.0);
+
+            let mut best_backend = String::new();
+            let mut best_latency = f64::MAX;
+
+            for (backend, variant) in &variants {
+                let matching = test_results
+                    .iter()
+                    .find(|r| r.backend == *backend && r.variant == *variant);
+                if let Some(r) = matching {
+                    let avg = if r.latency_us.is_empty() {
+                        0.0
+                    } else {
+                        metrics::mean(&r.latency_us)
+                    };
+                    let sd = metrics::stdev(&r.latency_us);
+                    let p95_val = metrics::p95(&r.latency_us) as f64;
+
+                    let cell = if use_ms {
+                        format!(
+                            "{:.2}\u{00b1}{:.2}ms (p95:{:.2}ms)",
+                            avg / 1000.0,
+                            sd / 1000.0,
+                            p95_val / 1000.0
+                        )
+                    } else {
+                        format!(
+                            "{:.1}\u{00b1}{:.1}\u{b5}s (p95:{:.1}\u{b5}s)",
+                            avg, sd, p95_val
+                        )
+                    };
+                    report.push_str(&format!(" {} |", cell));
+                    if avg > 0.0 && avg < best_latency {
+                        best_latency = avg;
+                        best_backend = format!("{} ({})", backend, variant);
+                    }
+                } else {
+                    report.push_str(" - |");
+                }
+            }
+            report.push_str(&format!(
+                " {} |\n",
+                if best_backend.is_empty() {
+                    "-"
+                } else {
+                    &best_backend
+                }
+            ));
+        }
+
+        report.push('\n');
+
+        let has_memory = category_results
+            .iter()
+            .any(|r| r.rss_kb > 0 || r.rss_peak_kb > 0);
+
+        if has_memory {
+            report.push_str("### Memory\n\n");
+            report.push_str("| Test | Variant | RSS | Peak RSS | Page Faults |\n");
+            report.push_str("|------|---------|-----|----------|-------------|\n");
+
+            let mut mem_rows: Vec<&BenchmarkResult> = category_results.to_vec();
+            mem_rows.sort_by(|a, b| {
+                (&a.test_name, &a.backend, &a.variant).cmp(&(&b.test_name, &b.backend, &b.variant))
+            });
+            for r in mem_rows {
+                let variant_label = format!("{} ({})", r.backend, r.variant);
+                report.push_str(&format!(
+                    "| {} | {} | {} | {} | {} |\n",
+                    r.test_name,
+                    variant_label,
+                    format_rss_mb(r.rss_kb),
+                    format_rss_mb(r.rss_peak_kb),
+                    format_faults(r.page_faults),
+                ));
+            }
+            report.push('\n');
+        }
+
+        type ExtraEntry = (String, f64);
+        type ExtraGroup = (String, String, Vec<ExtraEntry>);
+
+        let mut extras: Vec<ExtraGroup> = category_results
+            .iter()
+            .filter(|r| !r.extra.is_empty())
+            .map(|r| {
+                let mut pairs: Vec<(String, f64)> =
+                    r.extra.iter().map(|(k, &v)| (k.clone(), v)).collect();
+                pairs.sort_by(|a, b| a.0.cmp(&b.0));
+                (
+                    r.test_name.clone(),
+                    format!("{} ({})", r.backend, r.variant),
+                    pairs,
+                )
+            })
+            .collect();
+        extras.sort_by(|a, b| (&a.0, &a.1).cmp(&(&b.0, &b.1)));
+
+        if !extras.is_empty() {
+            report.push_str("### Extra Metrics\n\n");
+            report.push_str("| Test | Variant | Metric | Value |\n");
+            report.push_str("|------|---------|--------|-------|\n");
+
+            for (test_name, variant_label, pairs) in &extras {
+                for (key, val) in pairs {
+                    report.push_str(&format!(
+                        "| {} | {} | {} | {:.3} |\n",
+                        test_name, variant_label, key, val
+                    ));
+                }
+            }
+            report.push('\n');
+        }
+    }
+
+    report.push_str("## Summary\n\n");
+    report.push_str(&format!("- Total benchmarks: {}\n", results.len()));
+    report.push_str("- Categories: ");
+    report.push_str(&categories.keys().cloned().collect::<Vec<_>>().join(", "));
+    report.push('\n');
+
+    report
+}
+
+fn capitalize(s: &str) -> String {
+    let mut c = s.chars();
+    match c.next() {
+        None => String::new(),
+        Some(f) => f.to_uppercase().chain(c).collect(),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::collections::HashMap;
+
+    fn make_result(
+        category: &str,
+        test_name: &str,
+        backend: &str,
+        variant: &str,
+        latency_us: Vec<u64>,
+    ) -> BenchmarkResult {
+        BenchmarkResult {
+            category: category.to_string(),
+            test_name: test_name.to_string(),
+            backend: backend.to_string(),
+            variant: variant.to_string(),
+            latency_us,
+            rss_kb: 0,
+            rss_peak_kb: 0,
+            page_faults: 0,
+            extra: HashMap::new(),
+        }
+    }
+
+    #[test]
+    fn report_ordering_independent_of_input_order() {
+        let set_a = vec![
+            make_result(
+                "sequential",
+                "read_1mb",
+                "pread",
+                "default",
+                vec![100, 110, 105],
+            ),
+            make_result(
+                "sequential",
+                "read_1mb",
+                "mmap",
+                "default",
+                vec![80, 85, 90],
+            ),
+            make_result(
+                "sequential",
+                "read_4kb",
+                "pread",
+                "default",
+                vec![10, 12, 11],
+            ),
+            make_result("sequential", "read_4kb", "mmap", "default", vec![8, 9, 7]),
+        ];
+
+        let set_b = vec![
+            make_result("sequential", "read_4kb", "mmap", "default", vec![8, 9, 7]),
+            make_result(
+                "sequential",
+                "read_1mb",
+                "mmap",
+                "default",
+                vec![80, 85, 90],
+            ),
+            make_result(
+                "sequential",
+                "read_4kb",
+                "pread",
+                "default",
+                vec![10, 12, 11],
+            ),
+            make_result(
+                "sequential",
+                "read_1mb",
+                "pread",
+                "default",
+                vec![100, 110, 105],
+            ),
+        ];
+
+        let report_a = format_report(&set_a);
+        let report_b = format_report(&set_b);
+        assert_eq!(
+            report_a, report_b,
+            "Reports must be identical regardless of input order"
+        );
+    }
+}

crates/bench/src/runner.rs

@@ -0,0 +1,225 @@
+use crate::metrics::MetricsCollector;
+use crate::types::BenchmarkResult;
+use std::path::PathBuf;
+use std::sync::atomic::{AtomicBool, Ordering};
+
+/// All recognized benchmark suite identifiers.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum Suite {
+    Startup,
+    Render,
+    Jump,
+    Memory,
+    Growth,
+    Rotation,
+    Concurrent,
+}
+
+impl Suite {
+    /// All valid suite identifiers, in execution order.
+    pub const ALL: &[Suite] = &[
+        Suite::Startup,
+        Suite::Render,
+        Suite::Jump,
+        Suite::Memory,
+        Suite::Growth,
+        Suite::Rotation,
+        Suite::Concurrent,
+    ];
+
+    #[must_use]
+    pub const fn as_str(self) -> &'static str {
+        match self {
+            Suite::Startup => "startup",
+            Suite::Render => "render",
+            Suite::Jump => "jump",
+            Suite::Memory => "memory",
+            Suite::Growth => "growth",
+            Suite::Rotation => "rotation",
+            Suite::Concurrent => "concurrent",
+        }
+    }
+}
+
+impl std::str::FromStr for Suite {
+    type Err = String;
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        Suite::ALL
+            .iter()
+            .find(|suite| suite.as_str() == s)
+            .copied()
+            .ok_or_else(|| {
+                let valid = Suite::ALL
+                    .iter()
+                    .map(|s| s.as_str())
+                    .collect::<Vec<_>>()
+                    .join(", ");
+                format!("invalid value '{s}' for '--suites': valid values are {valid}")
+            })
+    }
+}
+
+pub struct BenchConfig {
+    pub test_file: PathBuf,
+    pub quick_mode: bool,
+    pub suites: Option<Vec<Suite>>,
+}
+
+/// Track whether we have already warned about VmHWM reset failure.
+/// Prevents duplicate warnings across multiple suite runs.
+static HWM_WARNED: AtomicBool = AtomicBool::new(false);
+
+pub fn warn_reset_hwm() {
+    if let Err(e) = MetricsCollector::reset_vm_hwm() {
+        if HWM_WARNED.swap(true, Ordering::Relaxed) {
+            return;
+        }
+        if e.kind() == std::io::ErrorKind::PermissionDenied {
+            eprintln!(
+                "WARNING: Failed to reset VmHWM via /proc/self/clear_refs: {e}. \
+                 Memory peak values may be contaminated across benchmark suites. \
+                 Try running as root."
+            );
+        } else {
+            eprintln!(
+                "WARNING: Failed to reset VmHWM via /proc/self/clear_refs: {e}. \
+                 Memory peak values may be contaminated across benchmark suites."
+            );
+        }
+    }
+}
+
+#[cfg(test)]
+fn reset_hwm_warned() {
+    HWM_WARNED.store(false, Ordering::Relaxed);
+}
+
+pub fn run_all(config: &BenchConfig) -> Vec<BenchmarkResult> {
+    let mut results = Vec::new();
+
+    let should_run = |suite: Suite| -> bool {
+        match &config.suites {
+            Some(suites) => suites.contains(&suite),
+            None => true,
+        }
+    };
+
+    if should_run(Suite::Startup) {
+        warn_reset_hwm();
+        results.extend(crate::suites::startup::run(config));
+    }
+    if should_run(Suite::Render) {
+        warn_reset_hwm();
+        results.extend(crate::suites::render::run(config));
+    }
+    if should_run(Suite::Jump) {
+        warn_reset_hwm();
+        results.extend(crate::suites::jump::run(config));
+    }
+    if should_run(Suite::Memory) {
+        warn_reset_hwm();
+        results.extend(crate::suites::memory::run(config));
+    }
+    if should_run(Suite::Growth) {
+        warn_reset_hwm();
+        results.extend(crate::suites::growth::run(config));
+    }
+    if should_run(Suite::Rotation) {
+        warn_reset_hwm();
+        results.extend(crate::suites::rotation::run(config));
+    }
+    if should_run(Suite::Concurrent) {
+        warn_reset_hwm();
+        results.extend(crate::suites::concurrent::run(config));
+    }
+
+    results
+}
+
+#[cfg(test)]
+mod tests {
+    use super::Suite;
+    use std::str::FromStr;
+
+    #[test]
+    fn parse_all_valid_suites() {
+        let expected = [
+            ("startup", Suite::Startup),
+            ("render", Suite::Render),
+            ("jump", Suite::Jump),
+            ("memory", Suite::Memory),
+            ("growth", Suite::Growth),
+            ("rotation", Suite::Rotation),
+            ("concurrent", Suite::Concurrent),
+        ];
+        for (s, expected_suite) in expected {
+            assert_eq!(
+                Suite::from_str(s).unwrap(),
+                expected_suite,
+                "failed to parse '{s}'"
+            );
+        }
+    }
+
+    #[test]
+    fn misspelled_suite_returns_error() {
+        let err = Suite::from_str("startp").unwrap_err();
+        assert!(
+            err.contains("invalid value 'startp'"),
+            "error should mention the bad value: {err}"
+        );
+    }
+
+    #[test]
+    fn error_message_lists_all_valid_values() {
+        let err = Suite::from_str("bogus").unwrap_err();
+        for name in Suite::ALL.iter().map(|s| s.as_str()) {
+            assert!(
+                err.contains(name),
+                "error should list valid suite '{name}': {err}"
+            );
+        }
+    }
+
+    #[test]
+    fn mixed_valid_invalid_stops_at_first_error() {
+        // "startup" is valid, "zzz" is not — collect hits the first Err
+        let names = ["startup".to_string(), "zzz".to_string()];
+        let result: Result<Vec<Suite>, _> = names.iter().map(|s| s.parse()).collect();
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn hwm_warned_flag_prevents_reentry() {
+        use super::reset_hwm_warned;
+        use std::sync::atomic::Ordering;
+
+        reset_hwm_warned();
+        assert!(
+            !super::HWM_WARNED.load(Ordering::Relaxed),
+            "flag should be false after reset"
+        );
+
+        // Simulate the flag being set (as warn_reset_hwm would do on error)
+        super::HWM_WARNED.store(true, Ordering::Relaxed);
+
+        // swap should now return true (old value), indicating already warned
+        assert!(
+            super::HWM_WARNED.swap(true, Ordering::Relaxed),
+            "swap should return the previous value (true)"
+        );
+    }
+
+    #[test]
+    fn warn_reset_hwm_does_not_panic() {
+        use super::reset_hwm_warned;
+
+        reset_hwm_warned();
+        // Whether reset_vm_hwm succeeds or fails, warn_reset_hwm must not panic.
+        // Multiple calls must also be safe.
+        super::warn_reset_hwm();
+        super::warn_reset_hwm();
+        super::warn_reset_hwm();
+    }
+}

crates/bench/src/suites/concurrent.rs

@@ -0,0 +1,117 @@
+use std::collections::HashMap;
+
+use crate::FileReaderBackend;
+use crate::metrics::MetricsCollector;
+use crate::mmap_reader::{
+    MmapReaderPhaseAware, MmapReaderPlain, MmapReaderPopulate, MmapReaderRandom,
+    MmapReaderSequential,
+};
+use crate::pread_reader::{PreadReaderPlain, PreadReaderRandom, PreadReaderSequential};
+use crate::runner::BenchConfig;
+use crate::types::BenchmarkResult;
+
+pub fn run(config: &BenchConfig) -> Vec<BenchmarkResult> {
+    let mut results = Vec::new();
+
+    results.extend(bench_parallel_reads::<MmapReaderPlain>(
+        "mmap", "plain", config,
+    ));
+    results.extend(bench_parallel_reads::<MmapReaderSequential>(
+        "mmap",
+        "sequential",
+        config,
+    ));
+    results.extend(bench_parallel_reads::<MmapReaderRandom>(
+        "mmap", "random", config,
+    ));
+    results.extend(bench_parallel_reads::<MmapReaderPopulate>(
+        "mmap", "populate", config,
+    ));
+    results.extend(bench_parallel_reads::<MmapReaderPhaseAware>(
+        "mmap",
+        "phase_aware",
+        config,
+    ));
+    results.extend(bench_parallel_reads::<PreadReaderPlain>(
+        "pread", "plain", config,
+    ));
+    results.extend(bench_parallel_reads::<PreadReaderRandom>(
+        "pread", "random", config,
+    ));
+    results.extend(bench_parallel_reads::<PreadReaderSequential>(
+        "pread",
+        "sequential",
+        config,
+    ));
+
+    results
+}
+
+fn bench_parallel_reads<B: FileReaderBackend + Send + 'static>(
+    backend: &str,
+    variant: &str,
+    config: &BenchConfig,
+) -> Vec<BenchmarkResult> {
+    let path = config.test_file.clone();
+    let iterations = if config.quick_mode { 250 } else { 1000 };
+
+    let total_lines = {
+        let reader = B::open(&path).expect("Failed to open file for line count");
+        let count = reader.total_lines();
+        reader.close();
+        count
+    };
+
+    let overall_start = std::time::Instant::now();
+    let num_threads = 4usize;
+
+    let handles: Vec<_> = (0..num_threads)
+        .map(|thread_id| {
+            let path = path.clone();
+            std::thread::spawn(move || {
+                let reader = B::open(&path).expect("Failed to open file in thread");
+                let mut latencies = Vec::with_capacity(iterations);
+
+                for i in 0..iterations {
+                    let line_idx = (thread_id * iterations + i) % total_lines.max(1);
+                    let t = std::time::Instant::now();
+                    let _ = reader.get_line(line_idx);
+                    latencies.push(t.elapsed().as_micros() as u64);
+                }
+
+                reader.close();
+                latencies
+            })
+        })
+        .collect();
+
+    let thread_latencies: Vec<Vec<u64>> = handles
+        .into_iter()
+        .map(|h| h.join().expect("Thread panicked"))
+        .collect();
+
+    let total_elapsed = overall_start.elapsed();
+
+    let all_latencies: Vec<u64> = thread_latencies.into_iter().flatten().collect();
+
+    let rss = MetricsCollector::read_rss();
+    let faults = MetricsCollector::read_page_faults();
+
+    let mut extra = HashMap::new();
+    extra.insert("num_threads".into(), num_threads as f64);
+    extra.insert("iterations_per_thread".into(), iterations as f64);
+    extra.insert("total_time_us".into(), total_elapsed.as_micros() as f64);
+    extra.insert("total_lines".into(), total_lines as f64);
+
+    vec![BenchmarkResult {
+        category: "concurrent".into(),
+        test_name: "parallel_reads".into(),
+        backend: backend.into(),
+        variant: variant.into(),
+        latency_us: all_latencies,
+        rss_kb: rss.vm_rss_kb,
+        rss_peak_kb: rss.vm_hwm_kb,
+        page_faults: faults.minor_faults + faults.major_faults,
+        extra,
+    }]
+}

crates/bench/src/suites/growth.rs

@@ -0,0 +1,295 @@
+use std::collections::HashMap;
+
+use crate::FileReaderBackend;
+use crate::data_gen;
+use crate::metrics::MetricsCollector;
+use crate::mmap_reader::MmapReaderPlain;
+use crate::pread_reader::PreadReaderPlain;
+use crate::runner::BenchConfig;
+use crate::types::BenchmarkResult;
+
+pub fn run(config: &BenchConfig) -> Vec<BenchmarkResult> {
+    let mut results = Vec::new();
+
+    let dir = tempfile::tempdir().expect("Failed to create temp dir");
+    results.extend(bench_append_visibility_mmap(config, dir.path()));
+    results.extend(bench_append_visibility_pread(config, dir.path()));
+    results.extend(bench_remap_cost(config, dir.path()));
+    results.extend(bench_scroll_during_append(config, dir.path()));
+    results.extend(bench_high_frequency_append(config, dir.path()));
+
+    results
+}
+
+fn bench_append_visibility_mmap(
+    config: &BenchConfig,
+    dir: &std::path::Path,
+) -> Vec<BenchmarkResult> {
+    let path = data_gen::generate_growable_file(dir).expect("Failed to create growable file");
+    let append_count: usize = if config.quick_mode { 100 } else { 1000 };
+
+    let mut reader = MmapReaderPlain::open(&path).expect("Failed to open growable file");
+    let original_lines = reader.total_lines();
+    let original_size = reader.file_size();
+
+    data_gen::append_lines(&path, append_count).expect("Failed to append lines");
+
+    let new_metadata = std::fs::metadata(&path).expect("Failed to read metadata");
+    let new_size = new_metadata.len() as usize;
+
+    let remap_start = std::time::Instant::now();
+    reader.remap(new_size).expect("Failed to remap");
+    let remap_elapsed = remap_start.elapsed();
+
+    let new_line_bytes =
+        reader.read_range(original_size, (new_size - original_size as usize).min(256));
+    let visible = new_line_bytes.is_some();
+
+    let rss = MetricsCollector::read_rss();
+    let faults = MetricsCollector::read_page_faults();
+
+    let mut extra = HashMap::new();
+    extra.insert("original_lines".into(), original_lines as f64);
+    extra.insert("appended_lines".into(), append_count as f64);
+    extra.insert("new_bytes_visible".into(), visible as u64 as f64);
+    extra.insert("original_size".into(), original_size as f64);
+    extra.insert("new_size".into(), new_size as f64);
+    extra.insert("remap_us".into(), remap_elapsed.as_micros() as f64);
+
+    reader.close();
+
+    vec![BenchmarkResult {
+        category: "growth".into(),
+        test_name: "append_visibility_mmap".into(),
+        backend: "mmap".into(),
+        variant: "plain".into(),
+        latency_us: vec![remap_elapsed.as_micros() as u64],
+        rss_kb: rss.vm_rss_kb,
+        rss_peak_kb: rss.vm_hwm_kb,
+        page_faults: faults.minor_faults + faults.major_faults,
+        extra,
+    }]
+}
+
+fn bench_append_visibility_pread(
+    config: &BenchConfig,
+    dir: &std::path::Path,
+) -> Vec<BenchmarkResult> {
+    let sub_dir = dir.join("pread_growth");
+    let path = data_gen::generate_growable_file(&sub_dir).expect("Failed to create growable file");
+    let append_count: usize = if config.quick_mode { 100 } else { 1000 };
+
+    let reader = PreadReaderPlain::open(&path).expect("Failed to open growable file");
+    let original_lines = reader.total_lines();
+    reader.close();
+
+    data_gen::append_lines(&path, append_count).expect("Failed to append lines");
+
+    let reopen_start = std::time::Instant::now();
+    let new_reader = PreadReaderPlain::open(&path).expect("Failed to reopen file");
+    let reopen_elapsed = reopen_start.elapsed();
+
+    let new_lines = new_reader.total_lines();
+    let can_read_new = new_lines > original_lines;
+    if can_read_new {
+        let last_line = new_lines.saturating_sub(1);
+        let _ = new_reader.get_line(last_line);
+    }
+
+    let rss = MetricsCollector::read_rss();
+    let faults = MetricsCollector::read_page_faults();
+
+    let mut extra = HashMap::new();
+    extra.insert("original_lines".into(), original_lines as f64);
+    extra.insert("appended_lines".into(), append_count as f64);
+    extra.insert("new_total_lines".into(), new_lines as f64);
+    extra.insert("reopen_us".into(), reopen_elapsed.as_micros() as f64);
+
+    new_reader.close();
+
+    vec![BenchmarkResult {
+        category: "growth".into(),
+        test_name: "append_visibility_pread".into(),
+        backend: "pread".into(),
+        variant: "plain".into(),
+        latency_us: vec![reopen_elapsed.as_micros() as u64],
+        rss_kb: rss.vm_rss_kb,
+        rss_peak_kb: rss.vm_hwm_kb,
+        page_faults: faults.minor_faults + faults.major_faults,
+        extra,
+    }]
+}
+
+fn bench_remap_cost(config: &BenchConfig, dir: &std::path::Path) -> Vec<BenchmarkResult> {
+    let sub_dir = dir.join("remap_cost");
+    let append_count: usize = if config.quick_mode { 100 } else { 1000 };
+    let iterations: usize = if config.quick_mode { 5 } else { 20 };
+
+    let mut latencies = Vec::with_capacity(iterations);
+
+    for _ in 0..iterations {
+        let path = data_gen::generate_growable_file(&sub_dir).expect("Failed to create file");
+        let mut reader = MmapReaderPlain::open(&path).expect("Failed to open file");
+
+        data_gen::append_lines(&path, append_count).expect("Failed to append");
+
+        let new_size = std::fs::metadata(&path).expect("metadata").len() as usize;
+        let t = std::time::Instant::now();
+        reader.remap(new_size).expect("remap failed");
+        latencies.push(t.elapsed().as_micros() as u64);
+
+        reader.close();
+        let _ = std::fs::remove_file(&path);
+    }
+
+    let rss = MetricsCollector::read_rss();
+    let faults = MetricsCollector::read_page_faults();
+
+    let mut extra = HashMap::new();
+    extra.insert("appended_per_iter".into(), append_count as f64);
+    extra.insert("iterations".into(), iterations as f64);
+
+    vec![BenchmarkResult {
+        category: "growth".into(),
+        test_name: "remap_cost".into(),
+        backend: "mmap".into(),
+        variant: "plain".into(),
+        latency_us: latencies,
+        rss_kb: rss.vm_rss_kb,
+        rss_peak_kb: rss.vm_hwm_kb,
+        page_faults: faults.minor_faults + faults.major_faults,
+        extra,
+    }]
+}
+
+fn bench_scroll_during_append(config: &BenchConfig, dir: &std::path::Path) -> Vec<BenchmarkResult> {
+    let sub_dir = dir.join("scroll_append");
+    let path = data_gen::generate_growable_file(&sub_dir).expect("Failed to create growable file");
+    let duration_secs: u64 = if config.quick_mode { 2 } else { 10 };
+    let append_rate: usize = if config.quick_mode { 1000 } else { 10000 };
+
+    let bg_path = path.clone();
+    let bg_handle = std::thread::spawn(move || {
+        let batch_size = 100;
+        let batch_interval =
+            std::time::Duration::from_micros(1_000_000 / (append_rate / batch_size).max(1) as u64);
+        let start = std::time::Instant::now();
+        while start.elapsed().as_secs() < duration_secs {
+            data_gen::append_lines(&bg_path, batch_size).ok();
+            std::thread::sleep(batch_interval);
+        }
+    });
+
+    let mut reader = PreadReaderPlain::open(&path).expect("Failed to open file");
+    let initial_lines = reader.total_lines();
+    let mut frame_latencies = Vec::new();
+    let mut current_line = 0usize;
+    let mut refresh_count: u64 = 0;
+    let mut none_count: u64 = 0;
+    let scroll_start = std::time::Instant::now();
+    let mut last_refresh = std::time::Instant::now();
+    let refresh_interval = std::time::Duration::from_millis(250);
+
+    while scroll_start.elapsed().as_secs() < duration_secs {
+        if last_refresh.elapsed() >= refresh_interval {
+            reader.refresh_index().ok();
+            refresh_count += 1;
+            last_refresh = std::time::Instant::now();
+            continue;
+        }
+
+        if let Some(_line) = reader.get_line(current_line) {
+            let t = std::time::Instant::now();
+            frame_latencies.push(t.elapsed().as_micros() as u64);
+            current_line += 1;
+        } else {
+            none_count += 1;
+            std::thread::sleep(std::time::Duration::from_millis(1));
+        }
+    }
+
+    let max_line = current_line;
+    assert!(
+        max_line > initial_lines,
+        "benchmark never read past initial {initial_lines} lines (max={max_line})"
+    );
+
+    reader.close();
+    bg_handle.join().ok();
+
+    let rss = MetricsCollector::read_rss();
+    let faults = MetricsCollector::read_page_faults();
+
+    let mut extra = HashMap::new();
+    extra.insert("duration_secs".into(), duration_secs as f64);
+    extra.insert("append_rate_per_sec".into(), append_rate as f64);
+    extra.insert("frames_rendered".into(), frame_latencies.len() as f64);
+    extra.insert("refresh_count".into(), refresh_count as f64);
+    extra.insert("none_count".into(), none_count as f64);
+    extra.insert("initial_lines".into(), initial_lines as f64);
+    extra.insert("max_line_seen".into(), max_line as f64);
+
+    vec![BenchmarkResult {
+        category: "growth".into(),
+        test_name: "scroll_during_append".into(),
+        backend: "pread".into(),
+        variant: "plain".into(),
+        latency_us: frame_latencies,
+        rss_kb: rss.vm_rss_kb,
+        rss_peak_kb: rss.vm_hwm_kb,
+        page_faults: faults.minor_faults + faults.major_faults,
+        extra,
+    }]
+}
+
+fn bench_high_frequency_append(
+    config: &BenchConfig,
+    dir: &std::path::Path,
+) -> Vec<BenchmarkResult> {
+    let sub_dir = dir.join("high_freq_append");
+    let path = data_gen::generate_growable_file(&sub_dir).expect("Failed to create growable file");
+    let duration_secs: u64 = if config.quick_mode { 3 } else { 30 };
+    let append_rate: usize = if config.quick_mode { 1000 } else { 10000 };
+    let batch_size: usize = 100;
+    let batches_per_sec = append_rate / batch_size;
+    let total_batches = (duration_secs as usize * batches_per_sec).max(1);
+
+    let mut detect_latencies = Vec::with_capacity(total_batches);
+
+    for _ in 0..total_batches {
+        data_gen::append_lines(&path, batch_size).expect("Failed to append");
+
+        let t = std::time::Instant::now();
+        if let Ok(reader) = PreadReaderPlain::open(&path) {
+            let total = reader.total_lines();
+            let _ = reader.get_line(total.saturating_sub(1));
+            reader.close();
+        }
+        detect_latencies.push(t.elapsed().as_micros() as u64);
+
+        std::thread::sleep(std::time::Duration::from_micros(
+            1_000_000 / batches_per_sec as u64,
+        ));
+    }
+
+    let rss = MetricsCollector::read_rss();
+    let faults = MetricsCollector::read_page_faults();
+
+    let mut extra = HashMap::new();
+    extra.insert("duration_secs".into(), duration_secs as f64);
+    extra.insert("append_rate_per_sec".into(), append_rate as f64);
+    extra.insert("batch_size".into(), batch_size as f64);
+    extra.insert("total_batches".into(), total_batches as f64);
+
+    vec![BenchmarkResult {
+        category: "growth".into(),
+        test_name: "high_frequency_append".into(),
+        backend: "pread".into(),
+        variant: "plain".into(),
+        latency_us: detect_latencies,
+        rss_kb: rss.vm_rss_kb,
+        rss_peak_kb: rss.vm_hwm_kb,
+        page_faults: faults.minor_faults + faults.major_faults,
+        extra,
+    }]
+}

crates/bench/src/suites/jump.rs

@@ -0,0 +1,251 @@
+use std::collections::HashMap;
+
+use super::FRAME_LINES;
+use crate::FileReaderBackend;
+use crate::metrics::MetricsCollector;
+use crate::mmap_reader::{
+    MmapReaderPhaseAware, MmapReaderPlain, MmapReaderPopulate, MmapReaderRandom,
+    MmapReaderSequential,
+};
+use crate::pread_reader::{PreadReaderPlain, PreadReaderRandom, PreadReaderSequential};
+use crate::runner::BenchConfig;
+use crate::types::BenchmarkResult;
+
+pub fn run(config: &BenchConfig) -> Vec<BenchmarkResult> {
+    let mut results = Vec::new();
+
+    results.extend(bench_near_jump::<MmapReaderPlain>("mmap", "plain", config));
+    results.extend(bench_near_jump::<MmapReaderSequential>(
+        "mmap",
+        "sequential",
+        config,
+    ));
+    results.extend(bench_near_jump::<MmapReaderRandom>(
+        "mmap", "random", config,
+    ));
+    results.extend(bench_near_jump::<MmapReaderPopulate>(
+        "mmap", "populate", config,
+    ));
+    results.extend(bench_near_jump::<MmapReaderPhaseAware>(
+        "mmap",
+        "phase_aware",
+        config,
+    ));
+    results.extend(bench_near_jump::<PreadReaderPlain>(
+        "pread", "plain", config,
+    ));
+    results.extend(bench_near_jump::<PreadReaderRandom>(
+        "pread", "random", config,
+    ));
+    results.extend(bench_near_jump::<PreadReaderSequential>(
+        "pread",
+        "sequential",
+        config,
+    ));
+
+    results.extend(bench_far_jump::<MmapReaderPlain>("mmap", "plain", config));
+    results.extend(bench_far_jump::<MmapReaderSequential>(
+        "mmap",
+        "sequential",
+        config,
+    ));
+    results.extend(bench_far_jump::<MmapReaderRandom>("mmap", "random", config));
+    results.extend(bench_far_jump::<MmapReaderPopulate>(
+        "mmap", "populate", config,
+    ));
+    results.extend(bench_far_jump::<MmapReaderPhaseAware>(
+        "mmap",
+        "phase_aware",
+        config,
+    ));
+    results.extend(bench_far_jump::<PreadReaderPlain>("pread", "plain", config));
+    results.extend(bench_far_jump::<PreadReaderRandom>(
+        "pread", "random", config,
+    ));
+    results.extend(bench_far_jump::<PreadReaderSequential>(
+        "pread",
+        "sequential",
+        config,
+    ));
+
+    results.extend(bench_jump_end::<MmapReaderPlain>("mmap", "plain", config));
+    results.extend(bench_jump_end::<PreadReaderPlain>("pread", "plain", config));
+
+    results.extend(bench_reverse_scan::<MmapReaderPlain>(
+        "mmap", "plain", config,
+    ));
+    results.extend(bench_reverse_scan::<PreadReaderPlain>(
+        "pread", "plain", config,
+    ));
+
+    results
+}
+
+fn bench_near_jump<B: FileReaderBackend>(
+    backend: &str,
+    variant: &str,
+    config: &BenchConfig,
+) -> Vec<BenchmarkResult> {
+    let path = &config.test_file;
+    let reader = B::open(path).expect("Failed to open file");
+    let total = reader.total_lines();
+    let iterations: usize = if config.quick_mode { 10 } else { 100 };
+
+    let mut latencies = Vec::with_capacity(iterations);
+    let mut current = 0usize;
+
+    for _ in 0..iterations {
+        let target = (current + 15).min(total.saturating_sub(1));
+        let t = std::time::Instant::now();
+        let _ = reader.get_line(target);
+        latencies.push(t.elapsed().as_micros() as u64);
+        current = target;
+    }
+
+    let rss = MetricsCollector::read_rss();
+    let faults = MetricsCollector::read_page_faults();
+
+    reader.close();
+
+    vec![BenchmarkResult {
+        category: "jump".into(),
+        test_name: "near_jump".into(),
+        backend: backend.into(),
+        variant: variant.into(),
+        latency_us: latencies,
+        rss_kb: rss.vm_rss_kb,
+        rss_peak_kb: rss.vm_hwm_kb,
+        page_faults: faults.minor_faults + faults.major_faults,
+        extra: HashMap::new(),
+    }]
+}
+
+fn bench_far_jump<B: FileReaderBackend>(
+    backend: &str,
+    variant: &str,
+    config: &BenchConfig,
+) -> Vec<BenchmarkResult> {
+    let path = &config.test_file;
+    let reader = B::open(path).expect("Failed to open file");
+    let total = reader.total_lines();
+    let repetitions: usize = if config.quick_mode { 3 } else { 10 };
+
+    let fractions = [0.25, 0.50, 0.75];
+    let mut latencies = Vec::new();
+    let mut extra = HashMap::new();
+
+    for &frac in &fractions {
+        let target = ((total as f64 * frac) as usize).min(total.saturating_sub(1));
+        for _ in 0..repetitions {
+            let t = std::time::Instant::now();
+            let _ = reader.get_line(target);
+            latencies.push(t.elapsed().as_micros() as u64);
+        }
+    }
+
+    extra.insert("jump_positions".into(), 3.0);
+
+    let rss = MetricsCollector::read_rss();
+    let faults = MetricsCollector::read_page_faults();
+
+    reader.close();
+
+    vec![BenchmarkResult {
+        category: "jump".into(),
+        test_name: "far_jump".into(),
+        backend: backend.into(),
+        variant: variant.into(),
+        latency_us: latencies,
+        rss_kb: rss.vm_rss_kb,
+        rss_peak_kb: rss.vm_hwm_kb,
+        page_faults: faults.minor_faults + faults.major_faults,
+        extra,
+    }]
+}
+
+fn bench_jump_end<B: FileReaderBackend>(
+    backend: &str,
+    variant: &str,
+    config: &BenchConfig,
+) -> Vec<BenchmarkResult> {
+    let path = &config.test_file;
+    let reader = B::open(path).expect("Failed to open file");
+    let total = reader.total_lines();
+    let iterations: usize = if config.quick_mode { 5 } else { 10 };
+
+    let last_line = total.saturating_sub(1);
+    let mut latencies = Vec::with_capacity(iterations);
+
+    for _ in 0..iterations {
+        let t = std::time::Instant::now();
+        let _ = reader.get_line(last_line);
+        latencies.push(t.elapsed().as_micros() as u64);
+    }
+
+    let rss = MetricsCollector::read_rss();
+    let faults = MetricsCollector::read_page_faults();
+
+    let mut extra = HashMap::new();
+    extra.insert("last_line_idx".into(), last_line as f64);
+
+    reader.close();
+
+    vec![BenchmarkResult {
+        category: "jump".into(),
+        test_name: "jump_end".into(),
+        backend: backend.into(),
+        variant: variant.into(),
+        latency_us: latencies,
+        rss_kb: rss.vm_rss_kb,
+        rss_peak_kb: rss.vm_hwm_kb,
+        page_faults: faults.minor_faults + faults.major_faults,
+        extra,
+    }]
+}
+
+fn bench_reverse_scan<B: FileReaderBackend>(
+    backend: &str,
+    variant: &str,
+    config: &BenchConfig,
+) -> Vec<BenchmarkResult> {
+    let path = &config.test_file;
+    let reader = B::open(path).expect("Failed to open file");
+    let total = reader.total_lines();
+
+    if total <= FRAME_LINES {
+        reader.close();
+        return vec![];
+    }
+
+    let iterations: usize = if config.quick_mode { 5 } else { 10 };
+    let mut latencies = Vec::with_capacity(iterations * FRAME_LINES);
+
+    for _ in 0..iterations {
+        let start = total - FRAME_LINES;
+        for i in (start..total).rev() {
+            let t = std::time::Instant::now();
+            let _ = reader.get_line(i);
+            latencies.push(t.elapsed().as_micros() as u64);
+        }
+    }
+
+    let rss = MetricsCollector::read_rss();
+    let faults = MetricsCollector::read_page_faults();
+    reader.close();
+
+    let mut extra = HashMap::new();
+    extra.insert("lines_per_scan".into(), FRAME_LINES as f64);
+    extra.insert("iterations".into(), iterations as f64);
+
+    vec![BenchmarkResult {
+        category: "jump".into(),
+        test_name: "reverse_scan".into(),
+        backend: backend.into(),
+        variant: variant.into(),
+        latency_us: latencies,
+        rss_kb: rss.vm_rss_kb,
+        rss_peak_kb: rss.vm_hwm_kb,
+        page_faults: faults.minor_faults + faults.major_faults,
+        extra,
+    }]
+}

crates/bench/src/suites/memory.rs

@@ -0,0 +1,237 @@
+use std::collections::HashMap;
+
+use crate::FileReaderBackend;
+use crate::metrics::MetricsCollector;
+use crate::mmap_reader::{
+    MmapReaderPhaseAware, MmapReaderPlain, MmapReaderPopulate, MmapReaderRandom,
+    MmapReaderSequential,
+};
+use crate::pread_reader::{PreadReaderPlain, PreadReaderRandom, PreadReaderSequential};
+use crate::runner::BenchConfig;
+use crate::types::BenchmarkResult;
+
+pub fn run(config: &BenchConfig) -> Vec<BenchmarkResult> {
+    let mut results = Vec::new();
+
+    results.extend(bench_idle_rss::<MmapReaderPlain>("mmap", "plain", config));
+    results.extend(bench_idle_rss::<MmapReaderSequential>(
+        "mmap",
+        "sequential",
+        config,
+    ));
+    results.extend(bench_idle_rss::<MmapReaderRandom>("mmap", "random", config));
+    results.extend(bench_idle_rss::<MmapReaderPopulate>(
+        "mmap", "populate", config,
+    ));
+    results.extend(bench_idle_rss::<MmapReaderPhaseAware>(
+        "mmap",
+        "phase_aware",
+        config,
+    ));
+    results.extend(bench_idle_rss::<PreadReaderPlain>("pread", "plain", config));
+    results.extend(bench_idle_rss::<PreadReaderRandom>(
+        "pread", "random", config,
+    ));
+    results.extend(bench_idle_rss::<PreadReaderSequential>(
+        "pread",
+        "sequential",
+        config,
+    ));
+
+    results.extend(bench_scroll_rss::<MmapReaderPlain>("mmap", "plain", config));
+    results.extend(bench_scroll_rss::<PreadReaderPlain>(
+        "pread", "plain", config,
+    ));
+
+    results.extend(bench_jump_end_rss::<MmapReaderPlain>(
+        "mmap", "plain", config,
+    ));
+    results.extend(bench_jump_end_rss::<PreadReaderPlain>(
+        "pread", "plain", config,
+    ));
+
+    results.extend(bench_rss_reclaim::<MmapReaderPlain>(
+        "mmap", "plain", config,
+    ));
+    results.extend(bench_rss_reclaim::<PreadReaderPlain>(
+        "pread", "plain", config,
+    ));
+
+    results
+}
+
+fn bench_idle_rss<B: FileReaderBackend>(
+    backend: &str,
+    variant: &str,
+    config: &BenchConfig,
+) -> Vec<BenchmarkResult> {
+    let reader = B::open(&config.test_file).expect("Failed to open file");
+    let rss = MetricsCollector::read_rss();
+    let faults = MetricsCollector::read_page_faults();
+
+    let mut extra = HashMap::new();
+    extra.insert("total_lines".into(), reader.total_lines() as f64);
+    extra.insert(
+        "file_size_mb".into(),
+        reader.file_size() as f64 / (1024.0 * 1024.0),
+    );
+
+    reader.close();
+
+    vec![BenchmarkResult {
+        category: "memory".into(),
+        test_name: "idle_rss".into(),
+        backend: backend.into(),
+        variant: variant.into(),
+        latency_us: vec![],
+        rss_kb: rss.vm_rss_kb,
+        rss_peak_kb: rss.vm_hwm_kb,
+        page_faults: faults.minor_faults + faults.major_faults,
+        extra,
+    }]
+}
+
+fn bench_scroll_rss<B: FileReaderBackend>(
+    backend: &str,
+    variant: &str,
+    config: &BenchConfig,
+) -> Vec<BenchmarkResult> {
+    let reader = B::open(&config.test_file).expect("Failed to open file");
+    let total = reader.total_lines();
+    let sample_interval = 100_000;
+    let max_lines = if config.quick_mode { 100_000 } else { total };
+
+    let upper = max_lines.min(total);
+    let mut rss_samples = Vec::new();
+    let mut hwm_samples = Vec::new();
+    let mut lines_read = 0usize;
+
+    for i in (0..upper).step_by(sample_interval) {
+        if reader.get_line(i).is_some() {
+            lines_read += 1;
+        }
+        let rss = MetricsCollector::read_rss();
+        rss_samples.push(rss.vm_rss_kb);
+        hwm_samples.push(rss.vm_hwm_kb);
+    }
+
+    let final_rss = MetricsCollector::read_rss();
+    let faults = MetricsCollector::read_page_faults();
+
+    let mut extra = HashMap::new();
+    extra.insert("rss_samples_count".into(), rss_samples.len() as f64);
+    extra.insert(
+        "max_rss_kb".into(),
+        rss_samples.iter().copied().fold(0u64, u64::max) as f64,
+    );
+    extra.insert(
+        "max_hwm_kb".into(),
+        hwm_samples.iter().copied().fold(0u64, u64::max) as f64,
+    );
+    extra.insert("lines_read".into(), lines_read as f64);
+
+    reader.close();
+
+    vec![BenchmarkResult {
+        category: "memory".into(),
+        test_name: "scroll_rss".into(),
+        backend: backend.into(),
+        variant: variant.into(),
+        latency_us: vec![],
+        rss_kb: final_rss.vm_rss_kb,
+        rss_peak_kb: final_rss.vm_hwm_kb,
+        page_faults: faults.minor_faults + faults.major_faults,
+        extra,
+    }]
+}
+
+fn bench_jump_end_rss<B: FileReaderBackend>(
+    backend: &str,
+    variant: &str,
+    config: &BenchConfig,
+) -> Vec<BenchmarkResult> {
+    let reader = B::open(&config.test_file).expect("Failed to open file");
+    let total = reader.total_lines();
+    let last_line = total.saturating_sub(1);
+
+    let _ = reader.get_line(last_line);
+
+    let rss = MetricsCollector::read_rss();
+    let faults = MetricsCollector::read_page_faults();
+
+    let mut extra = HashMap::new();
+    extra.insert("last_line_idx".into(), last_line as f64);
+    extra.insert("total_lines".into(), total as f64);
+
+    reader.close();
+
+    vec![BenchmarkResult {
+        category: "memory".into(),
+        test_name: "jump_end_rss".into(),
+        backend: backend.into(),
+        variant: variant.into(),
+        latency_us: vec![],
+        rss_kb: rss.vm_rss_kb,
+        rss_peak_kb: rss.vm_hwm_kb,
+        page_faults: faults.minor_faults + faults.major_faults,
+        extra,
+    }]
+}
+
+fn bench_rss_reclaim<B: FileReaderBackend>(
+    backend: &str,
+    variant: &str,
+    config: &BenchConfig,
+) -> Vec<BenchmarkResult> {
+    let reader = B::open(&config.test_file).expect("Failed to open file");
+    let total = reader.total_lines();
+    let last_line = total.saturating_sub(1);
+
+    let _ = reader.get_line(last_line);
+
+    let wait_secs: u64 = if config.quick_mode { 5 } else { 30 };
+    let sample_interval: u64 = 5;
+    let num_samples = (wait_secs / sample_interval) as usize;
+
+    let mut rss_samples = Vec::with_capacity(num_samples);
+    let mut hwm_samples = Vec::with_capacity(num_samples);
+
+    for _ in 0..num_samples {
+        std::thread::sleep(std::time::Duration::from_secs(sample_interval));
+        let rss = MetricsCollector::read_rss();
+        rss_samples.push(rss.vm_rss_kb);
+        hwm_samples.push(rss.vm_hwm_kb);
+    }
+
+    let final_rss = MetricsCollector::read_rss();
+    let faults = MetricsCollector::read_page_faults();
+    reader.close();
+
+    let mut extra = HashMap::new();
+    extra.insert("wait_total_secs".into(), wait_secs as f64);
+    extra.insert("rss_samples".into(), rss_samples.len() as f64);
+    if let (Some(&first), Some(&last)) = (rss_samples.first(), rss_samples.last()) {
+        extra.insert("rss_first_kb".into(), first as f64);
+        extra.insert("rss_last_kb".into(), last as f64);
+        extra.insert(
+            "rss_change_pct".into(),
+            if first > 0 {
+                ((last as f64 - first as f64) / first as f64) * 100.0
+            } else {
+                0.0
+            },
+        );
+    }
+
+    vec![BenchmarkResult {
+        category: "memory".into(),
+        test_name: "rss_reclaim".into(),
+        backend: backend.into(),
+        variant: variant.into(),
+        latency_us: vec![],
+        rss_kb: final_rss.vm_rss_kb,
+        rss_peak_kb: final_rss.vm_hwm_kb,
+        page_faults: faults.minor_faults + faults.major_faults,
+        extra,
+    }]
+}

crates/bench/src/suites/mod.rs

@@ -0,0 +1,12 @@
+pub mod concurrent;
+pub mod growth;
+pub mod jump;
+pub mod memory;
+pub mod render;
+pub mod rotation;
+pub mod startup;
+
+/// Number of lines read per single-frame render benchmark.
+/// Also used as the scan window size for reverse-scan benchmarks.
+/// Shared across suites to ensure consistent workload sizing.
+pub(crate) const FRAME_LINES: usize = 35;

crates/bench/src/suites/render.rs

@@ -0,0 +1,325 @@
+use std::collections::HashMap;
+
+use super::FRAME_LINES;
+use crate::FileReaderBackend;
+use crate::metrics::MetricsCollector;
+use crate::mmap_reader::{
+    MmapReaderPhaseAware, MmapReaderPlain, MmapReaderPopulate, MmapReaderRandom,
+    MmapReaderSequential,
+};
+use crate::pread_reader::{PreadReaderPlain, PreadReaderRandom, PreadReaderSequential};
+use crate::runner::BenchConfig;
+use crate::types::BenchmarkResult;
+
+pub fn run(config: &BenchConfig) -> Vec<BenchmarkResult> {
+    let mut results = Vec::new();
+
+    results.extend(bench_single_frame::<MmapReaderPlain>(
+        "mmap", "plain", config,
+    ));
+    results.extend(bench_single_frame::<MmapReaderSequential>(
+        "mmap",
+        "sequential",
+        config,
+    ));
+    results.extend(bench_single_frame::<MmapReaderRandom>(
+        "mmap", "random", config,
+    ));
+    results.extend(bench_single_frame::<MmapReaderPopulate>(
+        "mmap", "populate", config,
+    ));
+    results.extend(bench_single_frame::<MmapReaderPhaseAware>(
+        "mmap",
+        "phase_aware",
+        config,
+    ));
+    results.extend(bench_single_frame::<PreadReaderPlain>(
+        "pread", "plain", config,
+    ));
+    results.extend(bench_single_frame::<PreadReaderRandom>(
+        "pread", "random", config,
+    ));
+    results.extend(bench_single_frame::<PreadReaderSequential>(
+        "pread",
+        "sequential",
+        config,
+    ));
+
+    results.extend(bench_continuous_scroll::<MmapReaderPlain>(
+        "mmap", "plain", config,
+    ));
+    results.extend(bench_continuous_scroll::<MmapReaderSequential>(
+        "mmap",
+        "sequential",
+        config,
+    ));
+    results.extend(bench_continuous_scroll::<MmapReaderRandom>(
+        "mmap", "random", config,
+    ));
+    results.extend(bench_continuous_scroll::<MmapReaderPopulate>(
+        "mmap", "populate", config,
+    ));
+    results.extend(bench_continuous_scroll::<MmapReaderPhaseAware>(
+        "mmap",
+        "phase_aware",
+        config,
+    ));
+    results.extend(bench_continuous_scroll::<PreadReaderPlain>(
+        "pread", "plain", config,
+    ));
+    results.extend(bench_continuous_scroll::<PreadReaderRandom>(
+        "pread", "random", config,
+    ));
+    results.extend(bench_continuous_scroll::<PreadReaderSequential>(
+        "pread",
+        "sequential",
+        config,
+    ));
+
+    results
+}
+
+fn bench_single_frame<B: FileReaderBackend>(
+    backend: &str,
+    variant: &str,
+    config: &BenchConfig,
+) -> Vec<BenchmarkResult> {
+    let path = &config.test_file;
+    let reader = B::open(path).expect("Failed to open file");
+    let total = reader.total_lines();
+    let mut results = Vec::new();
+
+    for (pos_name, start_line) in select_frame_positions(total) {
+        let mut latencies = Vec::with_capacity(FRAME_LINES);
+        for i in 0..FRAME_LINES {
+            let line_idx = start_line + i;
+            if line_idx >= total {
+                break;
+            }
+            let t = std::time::Instant::now();
+            let _ = reader.get_line(line_idx);
+            latencies.push(t.elapsed().as_micros() as u64);
+        }
+
+        let rss = MetricsCollector::read_rss();
+        let faults = MetricsCollector::read_page_faults();
+
+        results.push(BenchmarkResult {
+            category: "render".into(),
+            test_name: format!("single_frame_{pos_name}"),
+            backend: backend.into(),
+            variant: variant.into(),
+            latency_us: latencies,
+            rss_kb: rss.vm_rss_kb,
+            rss_peak_kb: rss.vm_hwm_kb,
+            page_faults: faults.minor_faults + faults.major_faults,
+            extra: HashMap::new(),
+        });
+    }
+
+    reader.close();
+    results
+}
+
+/// Select non-overlapping benchmark positions for single-frame tests.
+///
+/// Returns `(label, start_line)` pairs. Positions are chosen so that no
+/// two frames overlap, skipping any position that cannot fit without
+/// colliding with a previously selected range.
+pub(crate) fn select_frame_positions(total: usize) -> Vec<(&'static str, usize)> {
+    let frame = FRAME_LINES;
+    if total == 0 {
+        return vec![];
+    }
+
+    let mut positions: Vec<(&'static str, usize)> = vec![("head", 0)];
+
+    // Need at least 3 full frames to place head, middle, and tail without overlap.
+    if total >= 3 * frame {
+        let head_end = frame;
+        let tail_start = total - frame;
+        let gap = tail_start - head_end;
+        // Center the middle frame within the gap, ensuring it fits entirely
+        let middle_start = head_end + gap.saturating_sub(frame) / 2;
+        positions.push(("middle", middle_start));
+        positions.push(("tail", tail_start));
+    }
+
+    positions
+}
+
+fn bench_continuous_scroll<B: FileReaderBackend>(
+    backend: &str,
+    variant: &str,
+    config: &BenchConfig,
+) -> Vec<BenchmarkResult> {
+    let path = &config.test_file;
+    let reader = B::open(path).expect("Failed to open file");
+    let total = reader.total_lines();
+
+    let iterations = if config.quick_mode { 100 } else { 1000 };
+    let timeout = if config.quick_mode {
+        std::time::Duration::from_secs(1)
+    } else {
+        std::time::Duration::from_secs(10)
+    };
+
+    let mut latencies = Vec::with_capacity(iterations);
+    let start = std::time::Instant::now();
+
+    for i in 0..iterations {
+        if start.elapsed() > timeout {
+            break;
+        }
+        let line_idx = i % total.max(1);
+        let t = std::time::Instant::now();
+        let _ = reader.get_line(line_idx);
+        latencies.push(t.elapsed().as_micros() as u64);
+    }
+
+    let rss = MetricsCollector::read_rss();
+    let faults = MetricsCollector::read_page_faults();
+
+    let mut extra = HashMap::new();
+    extra.insert("total_lines_scrolled".into(), latencies.len() as f64);
+
+    reader.close();
+
+    vec![BenchmarkResult {
+        category: "render".into(),
+        test_name: "continuous_scroll".into(),
+        backend: backend.into(),
+        variant: variant.into(),
+        latency_us: latencies,
+        rss_kb: rss.vm_rss_kb,
+        rss_peak_kb: rss.vm_hwm_kb,
+        page_faults: faults.minor_faults + faults.major_faults,
+        extra,
+    }]
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn starts(pos: &[(&str, usize)]) -> Vec<usize> {
+        pos.iter().map(|&(_, s)| s).collect()
+    }
+
+    fn labels<'a>(pos: &[(&'a str, usize)]) -> Vec<&'a str> {
+        pos.iter().map(|&(l, _)| l).collect()
+    }
+
+    fn ranges_overlap(a_start: usize, b_start: usize) -> bool {
+        let a_end = a_start + FRAME_LINES;
+        let b_end = b_start + FRAME_LINES;
+        a_start < b_end && b_start < a_end
+    }
+
+    #[test]
+    fn zero_lines_no_positions() {
+        assert!(select_frame_positions(0).is_empty());
+    }
+
+    #[test]
+    fn one_line_head_only() {
+        let pos = select_frame_positions(1);
+        assert_eq!(labels(&pos), vec!["head"]);
+        assert_eq!(starts(&pos), vec![0]);
+    }
+
+    #[test]
+    fn at_frame_lines_head_only() {
+        let pos = select_frame_positions(FRAME_LINES);
+        assert_eq!(labels(&pos), vec!["head"]);
+        assert_eq!(starts(&pos), vec![0]);
+    }
+
+    #[test]
+    fn just_above_threshold_head_middle_tail() {
+        let total = 3 * FRAME_LINES;
+        let pos = select_frame_positions(total);
+        assert_eq!(labels(&pos), vec!["head", "middle", "tail"]);
+        assert_eq!(starts(&pos)[0], 0);
+        assert_eq!(*starts(&pos).last().unwrap(), total - FRAME_LINES);
+        // Verify no overlap
+        for i in 0..pos.len() {
+            for j in (i + 1)..pos.len() {
+                assert!(
+                    !ranges_overlap(pos[i].1, pos[j].1),
+                    "overlap: {:?} @ {} vs {:?} @ {} (total={})",
+                    pos[i].0,
+                    pos[i].1,
+                    pos[j].0,
+                    pos[j].1,
+                    total
+                );
+            }
+        }
+    }
+
+    #[test]
+    fn no_overlap_at_total_70() {
+        // total=70 < 3*35=105, so only head is returned
+        let pos = select_frame_positions(70);
+        assert_eq!(labels(&pos), vec!["head"]);
+    }
+
+    #[test]
+    fn no_overlap_at_total_104() {
+        let pos = select_frame_positions(104);
+        for i in 0..pos.len() {
+            for j in (i + 1)..pos.len() {
+                assert!(
+                    !ranges_overlap(pos[i].1, pos[j].1),
+                    "overlap at total=104: {:?} @ {} vs {:?} @ {}",
+                    pos[i].0,
+                    pos[i].1,
+                    pos[j].0,
+                    pos[j].1
+                );
+            }
+        }
+    }
+
+    #[test]
+    fn no_overlap_at_total_105() {
+        let pos = select_frame_positions(105);
+        for i in 0..pos.len() {
+            for j in (i + 1)..pos.len() {
+                assert!(
+                    !ranges_overlap(pos[i].1, pos[j].1),
+                    "overlap at total=105: {:?} @ {} vs {:?} @ {}",
+                    pos[i].0,
+                    pos[i].1,
+                    pos[j].0,
+                    pos[j].1
+                );
+            }
+        }
+    }
+
+    #[test]
+    fn middle_is_centered_between_head_and_tail() {
+        let total = 1000;
+        let pos = select_frame_positions(total);
+        assert_eq!(pos.len(), 3);
+
+        let (_, head_start) = pos[0];
+        let (_, middle_start) = pos[1];
+        let (_, tail_start) = pos[2];
+
+        assert_eq!(head_start, 0);
+        assert_eq!(tail_start, total - FRAME_LINES);
+
+        let head_end = head_start + FRAME_LINES;
+        let gap = tail_start - head_end;
+        assert_eq!(middle_start, head_end + gap.saturating_sub(FRAME_LINES) / 2);
+    }
+
+    #[test]
+    fn large_file_all_three_positions() {
+        let pos = select_frame_positions(1_000_000);
+        assert_eq!(labels(&pos), vec!["head", "middle", "tail"]);
+    }
+}

crates/bench/src/suites/rotation.rs

@@ -0,0 +1,170 @@
+use std::collections::HashMap;
+
+use crate::FileReaderBackend;
+use crate::data_gen;
+use crate::metrics::MetricsCollector;
+use crate::mmap_reader::{self, MmapReaderPlain};
+use crate::pread_reader::PreadReaderPlain;
+use crate::runner::BenchConfig;
+use crate::types::BenchmarkResult;
+
+pub fn run(config: &BenchConfig) -> Vec<BenchmarkResult> {
+    let mut results = Vec::new();
+
+    let dir = tempfile::tempdir().expect("Failed to create temp dir");
+    results.extend(bench_truncate_safety_mmap(config, dir.path()));
+    results.extend(bench_truncate_safety_pread(config, dir.path()));
+    results.extend(bench_rotation_detection(config, dir.path()));
+
+    results
+}
+
+fn bench_truncate_safety_mmap(
+    _config: &BenchConfig,
+    dir: &std::path::Path,
+) -> Vec<BenchmarkResult> {
+    let sub_dir = dir.join("trunc_mmap");
+    let iterations: usize = if _config.quick_mode { 3 } else { 10 };
+
+    let mut latencies = Vec::with_capacity(iterations);
+    let mut sigbus_detected = 0usize;
+
+    for _ in 0..iterations {
+        let path = data_gen::generate_growable_file(&sub_dir).expect("Failed to create file");
+
+        mmap_reader::reset_sigbus_flag();
+
+        let reader = MmapReaderPlain::open(&path).expect("Failed to open file");
+        let original_size = reader.file_size();
+
+        let truncate_size = original_size / 2;
+        data_gen::truncate_file(&path, truncate_size).expect("Failed to truncate");
+
+        let t = std::time::Instant::now();
+        let mid_offset = original_size as u64 / 2;
+        let _ = reader.read_range(mid_offset, 64);
+        latencies.push(t.elapsed().as_micros() as u64);
+
+        if mmap_reader::sigbus_flag() {
+            sigbus_detected += 1;
+        }
+        reader.close();
+    }
+
+    let rss = MetricsCollector::read_rss();
+    let faults = MetricsCollector::read_page_faults();
+
+    let mut extra = HashMap::new();
+    extra.insert("iterations".into(), iterations as f64);
+    extra.insert("sigbus_detected".into(), sigbus_detected as f64);
+    extra.insert("crashed".into(), 0.0);
+
+    vec![BenchmarkResult {
+        category: "rotation".into(),
+        test_name: "truncate_safety_mmap".into(),
+        backend: "mmap".into(),
+        variant: "plain".into(),
+        latency_us: latencies,
+        rss_kb: rss.vm_rss_kb,
+        rss_peak_kb: rss.vm_hwm_kb,
+        page_faults: faults.minor_faults + faults.major_faults,
+        extra,
+    }]
+}
+
+fn bench_truncate_safety_pread(
+    _config: &BenchConfig,
+    dir: &std::path::Path,
+) -> Vec<BenchmarkResult> {
+    let sub_dir = dir.join("trunc_pread");
+    let iterations: usize = if _config.quick_mode { 3 } else { 10 };
+
+    let mut latencies = Vec::with_capacity(iterations);
+    let mut error_count = 0usize;
+
+    for _ in 0..iterations {
+        let path = data_gen::generate_growable_file(&sub_dir).expect("Failed to create file");
+
+        let reader = PreadReaderPlain::open(&path).expect("Failed to open file");
+        let original_size = reader.file_size();
+
+        let truncate_size = original_size / 2;
+        data_gen::truncate_file(&path, truncate_size).expect("Failed to truncate");
+
+        let t = std::time::Instant::now();
+        let mid_offset = original_size as u64 / 2;
+        let result = reader.read_range(mid_offset, 64);
+        latencies.push(t.elapsed().as_micros() as u64);
+
+        if result.is_none() {
+            error_count += 1;
+        }
+        reader.close();
+    }
+
+    let rss = MetricsCollector::read_rss();
+    let faults = MetricsCollector::read_page_faults();
+
+    let mut extra = HashMap::new();
+    extra.insert("iterations".into(), iterations as f64);
+    extra.insert("errors_returned".into(), error_count as f64);
+    extra.insert("crashed".into(), 0.0);
+
+    vec![BenchmarkResult {
+        category: "rotation".into(),
+        test_name: "truncate_safety_pread".into(),
+        backend: "pread".into(),
+        variant: "plain".into(),
+        latency_us: latencies,
+        rss_kb: rss.vm_rss_kb,
+        rss_peak_kb: rss.vm_hwm_kb,
+        page_faults: faults.minor_faults + faults.major_faults,
+        extra,
+    }]
+}
+
+fn bench_rotation_detection(_config: &BenchConfig, dir: &std::path::Path) -> Vec<BenchmarkResult> {
+    let sub_dir = dir.join("rotation_detect");
+    let iterations: usize = if _config.quick_mode { 3 } else { 10 };
+
+    let mut latencies = Vec::with_capacity(iterations);
+    let mut detected_count = 0usize;
+
+    for _ in 0..iterations {
+        let path = data_gen::generate_growable_file(&sub_dir).expect("Failed to create file");
+        let original_inode = MetricsCollector::get_inode(&path).expect("Failed to get inode");
+
+        let _rotated = data_gen::rotate_file(&path).expect("Failed to rotate file");
+
+        let t = std::time::Instant::now();
+        let detected = MetricsCollector::detect_rotation(original_inode, &path);
+        latencies.push(t.elapsed().as_micros() as u64);
+
+        if detected {
+            detected_count += 1;
+        }
+
+        let _ = std::fs::remove_file(&path);
+        let rotated = sub_dir.join("growable.log.1");
+        let _ = std::fs::remove_file(&rotated);
+    }
+
+    let rss = MetricsCollector::read_rss();
+    let faults = MetricsCollector::read_page_faults();
+
+    let mut extra = HashMap::new();
+    extra.insert("iterations".into(), iterations as f64);
+    extra.insert("rotations_detected".into(), detected_count as f64);
+
+    vec![BenchmarkResult {
+        category: "rotation".into(),
+        test_name: "rotation_detection".into(),
+        backend: "both".into(),
+        variant: "plain".into(),
+        latency_us: latencies,
+        rss_kb: rss.vm_rss_kb,
+        rss_peak_kb: rss.vm_hwm_kb,
+        page_faults: faults.minor_faults + faults.major_faults,
+        extra,
+    }]
+}

crates/bench/src/suites/startup.rs

@@ -0,0 +1,143 @@
+use std::collections::HashMap;
+use std::path::Path;
+
+use crate::FileReaderBackend;
+use crate::metrics::MetricsCollector;
+use crate::mmap_reader::{
+    MmapReaderPhaseAware, MmapReaderPlain, MmapReaderPopulate, MmapReaderRandom,
+    MmapReaderSequential,
+};
+use crate::pread_reader::{PreadReaderPlain, PreadReaderRandom, PreadReaderSequential};
+use crate::runner::BenchConfig;
+use crate::types::BenchmarkResult;
+
+pub fn run(config: &BenchConfig) -> Vec<BenchmarkResult> {
+    let mut results = Vec::new();
+
+    results.extend(bench_hot_open::<MmapReaderPlain>("mmap", "plain", config));
+    results.extend(bench_hot_open::<MmapReaderSequential>(
+        "mmap",
+        "sequential",
+        config,
+    ));
+    results.extend(bench_hot_open::<MmapReaderRandom>("mmap", "random", config));
+    results.extend(bench_hot_open::<MmapReaderPopulate>(
+        "mmap", "populate", config,
+    ));
+    results.extend(bench_hot_open::<MmapReaderPhaseAware>(
+        "mmap",
+        "phase_aware",
+        config,
+    ));
+    results.extend(bench_hot_open::<PreadReaderPlain>("pread", "plain", config));
+    results.extend(bench_hot_open::<PreadReaderRandom>(
+        "pread", "random", config,
+    ));
+    results.extend(bench_hot_open::<PreadReaderSequential>(
+        "pread",
+        "sequential",
+        config,
+    ));
+
+    if !config.quick_mode {
+        match MetricsCollector::clear_file_cache(&config.test_file) {
+            Ok(()) => {
+                results.extend(bench_cold_open::<MmapReaderPlain>("mmap", "plain", config));
+                results.extend(bench_cold_open::<MmapReaderSequential>(
+                    "mmap",
+                    "sequential",
+                    config,
+                ));
+                results.extend(bench_cold_open::<MmapReaderRandom>(
+                    "mmap", "random", config,
+                ));
+                results.extend(bench_cold_open::<MmapReaderPopulate>(
+                    "mmap", "populate", config,
+                ));
+                results.extend(bench_cold_open::<MmapReaderPhaseAware>(
+                    "mmap",
+                    "phase_aware",
+                    config,
+                ));
+                results.extend(bench_cold_open::<PreadReaderPlain>(
+                    "pread", "plain", config,
+                ));
+                results.extend(bench_cold_open::<PreadReaderRandom>(
+                    "pread", "random", config,
+                ));
+                results.extend(bench_cold_open::<PreadReaderSequential>(
+                    "pread",
+                    "sequential",
+                    config,
+                ));
+            }
+            Err(e) => {
+                eprintln!(
+                    "WARNING: Failed to clear file cache; skipping cold startup benchmarks because results would not be cold: {e}"
+                );
+            }
+        }
+    }
+
+    results
+}
+
+fn bench_hot_open<B: FileReaderBackend>(
+    backend: &str,
+    variant: &str,
+    config: &BenchConfig,
+) -> Vec<BenchmarkResult> {
+    open_and_measure::<B>(backend, variant, &config.test_file, "hot_open")
+}
+
+fn bench_cold_open<B: FileReaderBackend>(
+    backend: &str,
+    variant: &str,
+    config: &BenchConfig,
+) -> Vec<BenchmarkResult> {
+    match MetricsCollector::clear_file_cache(&config.test_file) {
+        Ok(()) => open_and_measure::<B>(backend, variant, &config.test_file, "cold_open"),
+        Err(e) => {
+            eprintln!(
+                "WARNING: Failed to clear file cache for {backend}/{variant}; skipping cold benchmark: {e}"
+            );
+            Vec::new()
+        }
+    }
+}
+
+fn open_and_measure<B: FileReaderBackend>(
+    backend: &str,
+    variant: &str,
+    path: &Path,
+    test_name: &str,
+) -> Vec<BenchmarkResult> {
+    let start = std::time::Instant::now();
+    let reader = B::open(path).expect("Failed to open file");
+    let elapsed = start.elapsed();
+
+    let rss = MetricsCollector::read_rss();
+    let faults = MetricsCollector::read_page_faults();
+
+    let result = BenchmarkResult {
+        category: "startup".into(),
+        test_name: test_name.into(),
+        backend: backend.into(),
+        variant: variant.into(),
+        latency_us: vec![elapsed.as_micros() as u64],
+        rss_kb: rss.vm_rss_kb,
+        rss_peak_kb: rss.vm_hwm_kb,
+        page_faults: faults.minor_faults + faults.major_faults,
+        extra: {
+            let mut m = HashMap::new();
+            m.insert("total_lines".into(), reader.total_lines() as f64);
+            m.insert(
+                "file_size_mb".into(),
+                reader.file_size() as f64 / (1024.0 * 1024.0),
+            );
+            m
+        },
+    };
+    reader.close();
+    vec![result]
+}

crates/bench/src/types.rs

@@ -0,0 +1,13 @@
+use std::collections::HashMap;
+
+pub struct BenchmarkResult {
+    pub category: String,
+    pub test_name: String,
+    pub backend: String,
+    pub variant: String,
+    pub latency_us: Vec<u64>,
+    pub rss_kb: u64,
+    pub rss_peak_kb: u64,
+    pub page_faults: u64,
+    pub extra: HashMap<String, f64>,
+}

crates/core/Cargo.toml

@@ -17,7 +17,8 @@ memmap2.workspace = true
 directories.workspace = true
 xxhash-rust.workspace = true
 bincode.workspace = true
+unicode-width.workspace = true
+tempfile.workspace = true
 
 [dev-dependencies]
 insta.workspace = true
-tempfile.workspace = true

crates/core/src/io/file_reader.rs

@@ -9,6 +9,14 @@ use crate::io::index_cache::IndexCache;
 use crate::io::line_index::LineIndex;
 use std::path::{Path, PathBuf};
 
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum AppendStatus {
+    Unchanged,
+    Appended(u64),
+    /// File shrank; mmap and index rebuilt via reload().
+    Reloaded,
+}
+
 pub struct FileReader {
     path: PathBuf,
     mmap: Option<memmap2::Mmap>,
@@ -25,20 +33,23 @@ impl FileReader {
         } else {
             // SAFETY: 使用只读 Mmap（非 MmapMut），文件以只读方式打开。
             // memmap2 内部持有文件描述符，确保 mmap 期间文件不会被关闭。
-            //
+            let m =
-            // ⚠️ Known limitation (Phase 5): 如果文件在 mmap 期间被外部进程截断，
+                unsafe { memmap2::Mmap::map(&file) }.map_err(|e| CoreError::Mmap(e.to_string()))?;
-            // 访问截断区域的内存会触发 SIGBUS（致命信号，无法恢复）。
+
-            // FileWatcher Phase 将添加文件修改检测和 re-mmap 机制来处理此情况。
+            // Layer 3: mmap 后立即 stat 同一 fd，检测截断（TOCTOU 缓解，非安全证明）
-            // 在 Phase 5 中，假设打开的文件不会被外部修改。
+            let current_size = file.metadata()?.len();
-            Some(unsafe { memmap2::Mmap::map(&file) }.map_err(|e| CoreError::Mmap(e.to_string()))?)
+            if current_size < m.len() as u64 {
+                None
+            } else {
+                Some(m)
+            }
         };
 
-        let line_index = {
+        // 直接从 mmap 快照构建行索引，确保索引与数据来自同一内存映射，
-            let mut reader = std::io::BufReader::new(&file);
+        // 消除 mmap + BufReader 双读之间的 TOCTOU 竞态窗口。
-            LineIndex::from_reader(&mut reader).map_err(|e| CoreError::Io {
+        let line_index = match &mmap {
-                source: e,
+            Some(m) => LineIndex::from_bytes(m.as_ref()),
-                context: "building line index".into(),
+            None => LineIndex::from_bytes(&[]),
-            })?
         };
 
         Ok(Self {
@@ -76,9 +87,8 @@ impl FileReader {
         }
     }
 
-    /// Save the line index cache to disk.
     pub fn save_cache(&self) -> std::io::Result<()> {
-        IndexCache::save(&self.path, &self.line_index)?;
+        IndexCache::save_with_hash(&self.path, &self.line_index, self.data())?;
         Ok(())
     }
 
@@ -94,27 +104,35 @@ impl FileReader {
         self.mmap = None;
 
         if file_size > 0 {
-            self.mmap = Some(
+            let m =
-                unsafe { memmap2::Mmap::map(&file) }.map_err(|e| CoreError::Mmap(e.to_string()))?,
+                unsafe { memmap2::Mmap::map(&file) }.map_err(|e| CoreError::Mmap(e.to_string()))?;
-            );
+            let current_size = file.metadata()?.len();
+            self.mmap = if current_size < m.len() as u64 {
+                None
+            } else {
+                Some(m)
+            };
         }
 
-        let mut reader = std::io::BufReader::new(&file);
+        self.line_index = match &self.mmap {
-        self.line_index = LineIndex::from_reader(&mut reader).map_err(|e| CoreError::Io {
+            Some(m) => LineIndex::from_bytes(m.as_ref()),
-            source: e,
+            None => LineIndex::from_bytes(&[]),
-            context: "rebuilding line index on reload".into(),
+        };
-        })?;
 
         Ok(())
     }
 
-    pub fn update_for_append(&mut self) -> Result<u64> {
+    pub fn update_for_append(&mut self) -> Result<AppendStatus> {
         let file = std::fs::File::open(&self.path)?;
         let new_size = file.metadata()?.len();
         let old_size = self.mmap.as_ref().map_or(0u64, |m| m.len() as u64);
 
-        if new_size <= old_size {
+        if new_size < old_size {
-            return Ok(0);
+            self.reload()?;
+            return Ok(AppendStatus::Reloaded);
+        }
+        if new_size == old_size {
+            return Ok(AppendStatus::Unchanged);
         }
 
         let old_lines = self.line_index.line_count() as u64;
@@ -123,11 +141,19 @@ impl FileReader {
         let mmap =
             unsafe { memmap2::Mmap::map(&file) }.map_err(|e| CoreError::Mmap(e.to_string()))?;
 
+        let current_size = file.metadata()?.len();
+        if current_size < mmap.len() as u64 {
+            self.line_index = LineIndex::from_bytes(&[]);
+            return Ok(AppendStatus::Reloaded);
+        }
+
         self.line_index
             .extend_from_bytes(&mmap[old_size as usize..], old_size);
         self.mmap = Some(mmap);
 
-        Ok(self.line_index.line_count() as u64 - old_lines)
+        Ok(AppendStatus::Appended(
+            self.line_index.line_count() as u64 - old_lines,
+        ))
     }
 }
 
@@ -305,8 +331,8 @@ mod tests {
             file.write_all(b"ccc\nddd\n").unwrap();
         }
 
-        let new_lines = reader.update_for_append().unwrap();
+        let status = reader.update_for_append().unwrap();
-        assert_eq!(new_lines, 2);
+        assert_eq!(status, AppendStatus::Appended(2));
         assert_eq!(reader.line_count(), 4);
         assert_eq!(reader.get_line(0), Some("aaa"));
         assert_eq!(reader.get_line(1), Some("bbb"));
@@ -320,8 +346,8 @@ mod tests {
         let mut reader = FileReader::open(f.path()).unwrap();
         assert_eq!(reader.line_count(), 1);
 
-        let new_lines = reader.update_for_append().unwrap();
+        let status = reader.update_for_append().unwrap();
-        assert_eq!(new_lines, 0);
+        assert_eq!(status, AppendStatus::Unchanged);
         assert_eq!(reader.line_count(), 1);
     }
 
@@ -341,8 +367,11 @@ mod tests {
             file.write_all(b"x\n").unwrap();
         }
 
-        let new_lines = reader.update_for_append().unwrap();
+        let status = reader.update_for_append().unwrap();
-        assert_eq!(new_lines, 0);
+        assert_eq!(status, AppendStatus::Reloaded);
+        assert_eq!(reader.line_count(), 1);
+        assert_eq!(reader.file_size(), 2);
+        assert_eq!(reader.get_line(0), Some("x"));
     }
 
     #[test]
@@ -368,4 +397,164 @@ mod tests {
         let idx = reader.line_index();
         assert_eq!(idx.line_count(), 2);
     }
+
+    // ─── TOCTOU 修复验证 ─────────────────────────────────────────────────
+    // open() 和 reload() 现在从 mmap 快照构建索引（from_bytes），
+    // 而非从独立 BufReader 读取（from_reader）。以下测试验证：
+    // 1. from_bytes 与 from_reader 产出完全一致的索引
+    // 2. 修改文件后 reload 仍返回正确内容（无残留旧状态）
+    // 3. 极端情况下（空文件、单行、跨块）open+reload 一致性
+
+    #[test]
+    fn test_open_from_bytes_matches_from_reader() {
+        let cases: Vec<&[u8]> = vec![
+            b"",
+            b"hello",
+            b"hello\n",
+            b"aaa\nbbb\nccc",
+            b"\n",
+            b"a\n\nb",
+        ];
+
+        let mut large_cases = Vec::new();
+        for &n in &[256usize, 300, 512] {
+            let mut buf = Vec::new();
+            for i in 0..n {
+                buf.extend_from_slice(format!("line{}\n", i).as_bytes());
+            }
+            large_cases.push(buf);
+        }
+
+        let all_contents: Vec<&[u8]> = cases
+            .iter()
+            .copied()
+            .chain(large_cases.iter().map(Vec::as_slice))
+            .collect();
+
+        for content in all_contents {
+            let f = create_temp_file(content);
+            let reader = FileReader::open(f.path()).unwrap();
+
+            let from_bytes_idx = LineIndex::from_bytes(content);
+            let mut buf_reader = std::io::BufReader::new(content);
+            let from_reader_idx = LineIndex::from_reader(&mut buf_reader).unwrap();
+
+            assert_eq!(
+                reader.line_index().total_lines(),
+                from_bytes_idx.total_lines(),
+                "open() line_count should match from_bytes for {:?}B file",
+                content.len()
+            );
+            assert_eq!(
+                from_bytes_idx.total_lines(),
+                from_reader_idx.total_lines(),
+                "from_bytes should match from_reader total_lines for {:?}B file",
+                content.len()
+            );
+            assert_eq!(
+                from_bytes_idx.sampled_offsets(),
+                from_reader_idx.sampled_offsets(),
+                "from_bytes should match from_reader sampled_offsets for {:?}B file",
+                content.len()
+            );
+            assert_eq!(
+                from_bytes_idx.has_trailing_newline(),
+                from_reader_idx.has_trailing_newline(),
+                "from_bytes should match from_reader has_trailing_newline for {:?}B file",
+                content.len()
+            );
+        }
+    }
+
+    #[test]
+    fn test_reload_after_external_modify_returns_correct_content() {
+        let f = create_temp_file(b"aaa\nbbb\n");
+        let mut reader = FileReader::open(f.path()).unwrap();
+        assert_eq!(reader.get_line(0), Some("aaa"));
+        assert_eq!(reader.get_line(1), Some("bbb"));
+        assert_eq!(reader.line_count(), 2);
+
+        // 外部修改：覆盖写入新内容
+        {
+            use std::io::Write;
+            std::fs::write(f.path(), b"xxx\nyyy\nzzz\n").unwrap();
+        }
+
+        reader.reload().unwrap();
+        assert_eq!(reader.line_count(), 3);
+        assert_eq!(reader.get_line(0), Some("xxx"));
+        assert_eq!(reader.get_line(1), Some("yyy"));
+        assert_eq!(reader.get_line(2), Some("zzz"));
+    }
+
+    #[test]
+    fn test_reload_after_truncate_then_rewrite_no_stale_data() {
+        let f = create_temp_file(b"line0\nline1\nline2\nline3\n");
+        let mut reader = FileReader::open(f.path()).unwrap();
+        assert_eq!(reader.line_count(), 4);
+
+        // 截断后写入更短内容
+        {
+            use std::io::Write;
+            let mut file = std::fs::OpenOptions::new()
+                .write(true)
+                .truncate(true)
+                .open(f.path())
+                .unwrap();
+            file.write_all(b"new\n").unwrap();
+        }
+
+        reader.reload().unwrap();
+        assert_eq!(reader.line_count(), 1);
+        assert_eq!(reader.get_line(0), Some("new"));
+        assert_eq!(
+            reader.get_line(1),
+            None,
+            "old line should not be accessible"
+        );
+    }
+
+    #[test]
+    fn test_open_reload_idempotent_cross_block() {
+        let mut content = Vec::new();
+        for i in 0..600u32 {
+            content.extend_from_slice(format!("line{}\n", i).as_bytes());
+        }
+        let f = create_temp_file(&content);
+
+        let reader = FileReader::open(f.path()).unwrap();
+        let mut reloaded = FileReader::open(f.path()).unwrap();
+        reloaded.reload().unwrap();
+
+        assert_eq!(reader.line_count(), reloaded.line_count());
+        for i in 0..600 {
+            assert_eq!(
+                reader.get_line(i),
+                reloaded.get_line(i),
+                "line {i} mismatch between open and reload"
+            );
+        }
+    }
+
+    #[test]
+    fn test_open_stat_after_mmap_detects_truncation() {
+        let content = b"line0\nline1\nline2\nline3\n";
+        let f = create_temp_file(content);
+
+        let reader = FileReader::open(f.path()).unwrap();
+        assert_eq!(reader.line_count(), 4);
+
+        {
+            use std::io::Write;
+            let _ = std::fs::OpenOptions::new()
+                .write(true)
+                .truncate(true)
+                .open(f.path())
+                .unwrap();
+        }
+
+        let reader = FileReader::open(f.path()).unwrap();
+        assert_eq!(reader.line_count(), 0);
+        assert_eq!(reader.file_size(), 0);
+    }
 }

crates/core/src/io/index_cache.rs

@@ -1,19 +1,37 @@
 use std::io::{Read as _, Write as _};
 use std::path::Path;
 
-use crate::io::cache_util::{cache_path, CACHE_VERSION};
+use crate::io::cache_util::{CACHE_VERSION, cache_path};
 use crate::io::line_index::LineIndex;
 
 pub struct IndexCache;
 
-impl IndexCache {
+/// Write `buf` to `dest` atomically using a unique temporary file in the same directory.
-    /// Save a `LineIndex` to disk using atomic write (write to .tmp, then rename).
+///
-    pub fn save(file_path: &Path, index: &LineIndex) -> std::io::Result<()> {
+/// Each call creates its own temp file via `tempfile::Builder`, eliminating collisions
-        let dest = cache_path(file_path).ok_or_else(|| {
+/// when multiple threads (or processes) save to the same cache path concurrently.
-            std::io::Error::new(std::io::ErrorKind::NotFound, "cannot determine cache path")
+/// The temp file is created in `dest.parent()` so the final `rename` stays on the
+/// same filesystem and remains atomic.
+fn write_cache_atomically(dest: &Path, buf: &[u8]) -> std::io::Result<()> {
+    let dir = dest.parent().ok_or_else(|| {
+        std::io::Error::new(
+            std::io::ErrorKind::InvalidInput,
+            "cache destination has no parent directory",
+        )
     })?;
 
-        let file_hash = compute_file_hash(file_path)?;
+    let mut tmp = tempfile::Builder::new()
+        .prefix("index-cache-")
+        .suffix(".tmp")
+        .tempfile_in(dir)?;
+
+    tmp.as_file_mut().write_all(buf)?;
+    tmp.as_file_mut().sync_all()?;
+
+    tmp.persist(dest).map(|_| ()).map_err(|e| e.error)
+}
+
+fn encode_cache(file_hash: u64, index: &LineIndex) -> std::io::Result<Vec<u8>> {
     let index_bytes = bincode::serialize(index)
         .map_err(|e| std::io::Error::new(std::io::ErrorKind::InvalidData, e.to_string()))?;
 
@@ -21,16 +39,36 @@ impl IndexCache {
     buf.push(CACHE_VERSION);
     buf.extend_from_slice(&file_hash.to_le_bytes());
     buf.extend_from_slice(&index_bytes);
+    Ok(buf)
+}
 
-        let tmp_path = dest.with_extension("index.tmp");
+impl IndexCache {
-        {
+    /// Save a `LineIndex` to disk using atomic write (unique temp file, then rename).
-            let mut f = std::fs::File::create(&tmp_path)?;
+    ///
-            f.write_all(&buf)?;
+    /// The file hash is derived from `data` (the same byte slice used to build the index),
-            f.sync_all()?;
+    /// avoiding TOCTOU issues from re-reading the file from disk.
+    pub fn save_with_hash(file_path: &Path, index: &LineIndex, data: &[u8]) -> std::io::Result<()> {
+        let dest = cache_path(file_path).ok_or_else(|| {
+            std::io::Error::new(std::io::ErrorKind::NotFound, "cannot determine cache path")
+        })?;
+
+        let file_hash = compute_data_hash(data);
+        let buf = encode_cache(file_hash, index)?;
+        write_cache_atomically(&dest, &buf)
     }
-        std::fs::rename(&tmp_path, &dest)?;
 
-        Ok(())
+    /// Save a `LineIndex` to disk, computing the hash by re-reading the file.
+    ///
+    /// Prefer [`save_with_hash`] when the source data is already in memory
+    /// to avoid a TOCTOU race between indexing and hash computation.
+    pub fn save(file_path: &Path, index: &LineIndex) -> std::io::Result<()> {
+        let dest = cache_path(file_path).ok_or_else(|| {
+            std::io::Error::new(std::io::ErrorKind::NotFound, "cannot determine cache path")
+        })?;
+
+        let file_hash = compute_file_hash(file_path)?;
+        let buf = encode_cache(file_hash, index)?;
+        write_cache_atomically(&dest, &buf)
     }
 
     /// Load a cached `LineIndex` from disk.
@@ -60,6 +98,32 @@ impl IndexCache {
     }
 }
 
+/// Compute a fingerprint from in-memory data, using the same algorithm as `compute_file_hash`.
+/// Returns 0 for empty data.
+pub fn compute_data_hash(data: &[u8]) -> u64 {
+    let len = data.len();
+    if len == 0 {
+        return 0;
+    }
+
+    let head_size = 4096.min(len);
+    let tail_size = 4096.min(len);
+
+    let mut hasher_state = xxhash_rust::xxh3::Xxh3::new();
+    hasher_state.update(&data[..head_size]);
+
+    if len > head_size + tail_size {
+        hasher_state.update(&data[len - tail_size..]);
+    } else {
+        // For small files the tail overlaps with the head; hash from the real tail start.
+        let tail_start = len.saturating_sub(tail_size);
+        hasher_state.update(&data[tail_start..]);
+    }
+    hasher_state.update(&(len as u64).to_le_bytes());
+
+    hasher_state.digest()
+}
+
 /// Compute a fast fingerprint of the file: xxhash of (head 4KB + tail 4KB + file size).
 /// Returns 0 for empty files.
 fn compute_file_hash(file_path: &Path) -> std::io::Result<u64> {
@@ -243,4 +307,81 @@ mod tests {
 
         assert_ne!(h1, h2, "hash should change when content changes");
     }
+
+    #[test]
+    fn test_concurrent_save_with_hash_no_corruption() {
+        use std::sync::{Arc, Barrier};
+
+        let file = make_test_file(300);
+        let data = std::fs::read(file.path()).unwrap();
+
+        let num_threads = 8;
+        let iterations = 50;
+        let barrier = Arc::new(Barrier::new(num_threads));
+        let path = file.path().to_path_buf();
+
+        let handles: Vec<_> = (0..num_threads)
+            .map(|_| {
+                let barrier = Arc::clone(&barrier);
+                let path = path.clone();
+                let data = data.clone();
+                std::thread::spawn(move || {
+                    let index = LineIndex::from_bytes(&data);
+                    barrier.wait();
+                    for _ in 0..iterations {
+                        IndexCache::save_with_hash(&path, &index, &data)
+                            .expect("concurrent save_with_hash should succeed");
+                    }
+                })
+            })
+            .collect();
+
+        for h in handles {
+            h.join().expect("thread should not panic");
+        }
+
+        let loaded = IndexCache::load(file.path()).expect("final cache should load successfully");
+        let expected = LineIndex::from_bytes(&data);
+        assert_eq!(loaded.line_count(), expected.line_count());
+        assert_eq!(
+            loaded.sampled_offsets(),
+            expected.sampled_offsets(),
+            "concurrent writes must not produce interleaved data"
+        );
+    }
+
+    #[test]
+    fn test_concurrent_save_same_dest_all_succeed() {
+        use std::sync::{Arc, Barrier};
+
+        let dir = tempfile::tempdir().unwrap();
+        let dest = dir.path().join("test.index");
+
+        let payloads: Vec<Vec<u8>> = (0..8).map(|i| vec![i; 64 * 1024]).collect();
+        let barrier = Arc::new(Barrier::new(8));
+
+        let handles: Vec<_> = payloads
+            .into_iter()
+            .map(|payload| {
+                let barrier = Arc::clone(&barrier);
+                let dest = dest.clone();
+                std::thread::spawn(move || {
+                    barrier.wait();
+                    write_cache_atomically(&dest, &payload)
+                        .expect("concurrent atomic write should succeed");
+                })
+            })
+            .collect();
+
+        for h in handles {
+            h.join().expect("thread should not panic");
+        }
+
+        let final_data = std::fs::read(&dest).expect("dest file should exist");
+        assert_eq!(
+            final_data.len(),
+            64 * 1024,
+            "final file must be exactly one payload"
+        );
+    }
 }

crates/core/src/io/line_index.rs

@@ -20,14 +20,14 @@ const BLOCK_SIZE: usize = 256;
 pub struct LineIndex {
     // 采样偏移量：每 BLOCK_SIZE 行记录一个起始字节偏移。
     // sampled_offsets[i] 存储第 (i * BLOCK_SIZE) 行的字节起始位置。
-    sampled_offsets: Vec<u64>,
+    pub(crate) sampled_offsets: Vec<u64>,
 
     // 文件总行数。
-    total_lines: u64,
+    pub(crate) total_lines: u64,
 
     // 文件最后一个字节是否是换行符 \n。
     #[allow(dead_code)]
-    has_trailing_newline: bool,
+    pub(crate) has_trailing_newline: bool,
 }
 
 impl LineIndex {
@@ -158,7 +158,7 @@ impl LineIndex {
 
         // If the junction falls on a block boundary, record the start offset
         // (analogous to from_bytes always pushing offset 0 for line 0).
-        if starts_new_line && (old_total as usize) % BLOCK_SIZE == 0 {
+        if starts_new_line && (old_total as usize).is_multiple_of(BLOCK_SIZE) {
             self.sampled_offsets.push(start_offset);
         }
 
@@ -212,15 +212,17 @@ impl LineIndex {
         self.total_lines as usize
     }
 
-    // ─── getter 方法 ────────────────────────────────────────────────────
+    #[cfg(test)]
     pub(crate) fn sampled_offsets(&self) -> &[u64] {
         &self.sampled_offsets
     }
 
+    #[cfg(test)]
     pub(crate) fn total_lines(&self) -> u64 {
         self.total_lines
     }
 
+    #[cfg(test)]
     pub(crate) fn has_trailing_newline(&self) -> bool {
         self.has_trailing_newline
     }
@@ -233,7 +235,7 @@ impl LineIndex {
         }
         let block = idx / BLOCK_SIZE;
         let offset_in_block = idx % BLOCK_SIZE;
-        let mut pos = self.sampled_offsets[block] as usize;
+        let mut pos = (*self.sampled_offsets.get(block)?) as usize;
         for _ in 0..offset_in_block {
             match memchr::memchr(b'\n', &data[pos..]) {
                 Some(rel) => pos = pos + rel + 1,

crates/core/src/io/mod.rs

@@ -18,4 +18,5 @@ pub mod cache_util;
 pub mod index_cache;
 pub mod line_sampler;
 pub mod progressive_reader;
+pub mod read_cache;
 pub mod wrap;

crates/core/src/io/progressive_reader.rs

@@ -1,13 +1,32 @@
 use std::cell::RefCell;
 use std::fmt;
+use std::io::BufRead;
 use std::path::{Path, PathBuf};
 
 use crate::error::{CoreError, Result};
-use crate::io::file_reader::FileReader;
+use crate::io::file_reader::{AppendStatus, FileReader};
 use crate::io::index_cache::IndexCache;
 use crate::io::line_index::LineIndex;
 use crate::io::line_sampler::sample_line_count;
-use crate::io::wrap::{format_json_line, wrap_line_chars, MAX_WRAP_INPUT_LEN};
+use crate::io::wrap::{MAX_WRAP_INPUT_LEN, format_json_line};
+
+// ─── Cancel-aware channel helpers ────────────────────────────────────────────
+
+/// Send a message on `tx`, but abort if `cancel_rx` fires first.
+///
+/// Uses `crossbeam_channel::select!` so the thread sleeps efficiently instead
+/// of busy-looping.  Used for terminal messages (Complete / Error) that must
+/// not be silently dropped while the receiver is still alive.
+fn send_cancelable<T>(
+    tx: &crossbeam_channel::Sender<T>,
+    msg: T,
+    cancel_rx: &crossbeam_channel::Receiver<()>,
+) {
+    crossbeam_channel::select! {
+        send(tx, msg) -> _ => {}
+        recv(cancel_rx) -> _ => {}
+    }
+}
 
 // ─── IndexerMessage ──────────────────────────────────────────────────────────
 
@@ -101,7 +120,7 @@ impl VisualHeightIndex {
         }
     }
 
-    fn with_params(mut self, json_format: bool, terminal_width: usize) -> Self {
+    pub fn with_params(mut self, json_format: bool, terminal_width: usize) -> Self {
         self.json_format = json_format;
         self.terminal_width = terminal_width;
         self
@@ -158,6 +177,31 @@ impl VisualHeightIndex {
             self.total_visual_rows += h as u64;
         }
     }
+
+    /// Replace the visual height of the last logical line. O(1).
+    ///
+    /// Must be called **before** `extend_from_heights` so that the last line
+    /// index still refers to the pre-extension line.
+    pub fn replace_last_line_height(&mut self, new_height: usize) {
+        let n = self.prefix_sums.len();
+        if n < 2 {
+            return;
+        }
+        let last_line = n - 2;
+        let old_height = self.prefix_sums[last_line + 1] - self.prefix_sums[last_line];
+        let new_height = new_height as u64;
+        if new_height == old_height {
+            return;
+        }
+        let delta = new_height.abs_diff(old_height);
+        if new_height > old_height {
+            self.prefix_sums[last_line + 1] += delta;
+            self.total_visual_rows += delta;
+        } else {
+            self.prefix_sums[last_line + 1] -= delta;
+            self.total_visual_rows -= delta;
+        }
+    }
 }
 
 // ─── VisualHeightRebuildResult ────────────────────────────────────────────────
@@ -179,6 +223,9 @@ pub fn compute_line_visual_height(
     }
     if json_format {
         let formatted = format_json_line(line_text);
+        if formatted.len() > MAX_WRAP_INPUT_LEN {
+            return 1;
+        }
         compute_text_visual_height(&formatted, terminal_width)
     } else {
         compute_text_visual_height(line_text, terminal_width)
@@ -188,29 +235,11 @@ pub fn compute_line_visual_height(
 fn compute_text_visual_height(text: &str, width: usize) -> usize {
     let mut height = 0;
     for sub_line in text.split('\n') {
-        height += wrap_line_chars(sub_line, width).len();
+        height += crate::io::wrap::wrap_line_count(sub_line, width);
     }
     height.max(1)
 }
 
-fn compute_visual_heights(
-    reader: &FileReader,
-    terminal_width: usize,
-    json_format: bool,
-) -> Vec<usize> {
-    let line_count = reader.line_count();
-    let mut visual_heights = Vec::with_capacity(line_count);
-    for i in 0..line_count {
-        let line_text = reader.get_line(i).unwrap_or("");
-        visual_heights.push(compute_line_visual_height(
-            line_text,
-            terminal_width,
-            json_format,
-        ));
-    }
-    visual_heights
-}
-
 // ─── ReaderState ─────────────────────────────────────────────────────────────
 
 pub enum ReaderState {
@@ -234,8 +263,8 @@ pub enum ReaderState {
 pub fn spawn_indexer(
     path: PathBuf,
     generation: u64,
-    terminal_width: usize,
+    _terminal_width: usize,
-    json_format: bool,
+    _json_format: bool,
     cancel_rx: crossbeam_channel::Receiver<()>,
 ) -> crossbeam_channel::Receiver<IndexerMessage> {
     let (tx, rx) = crossbeam_channel::bounded(10);
@@ -244,62 +273,84 @@ pub fn spawn_indexer(
         let file = match std::fs::File::open(&path) {
             Ok(f) => f,
             Err(e) => {
-                let _ = tx.send(IndexerMessage::Error {
+                send_cancelable(
+                    &tx,
+                    IndexerMessage::Error {
                         generation,
                         message: e.to_string(),
-                });
+                    },
+                    &cancel_rx,
+                );
                 return;
             }
         };
-        let file_size = match file.metadata() {
+        let target_len = match file.metadata() {
             Ok(m) => m.len(),
             Err(e) => {
-                let _ = tx.send(IndexerMessage::Error {
+                send_cancelable(
+                    &tx,
+                    IndexerMessage::Error {
                         generation,
                         message: e.to_string(),
-                });
+                    },
+                    &cancel_rx,
+                );
                 return;
             }
         };
 
-        let mmap = if file_size == 0 {
+        let mut buf_reader = std::io::BufReader::with_capacity(64 * 1024, file);
-            None
+        let mut sampled_offsets: Vec<u64> = vec![0];
-        } else {
+        let mut next_line_idx: usize = 1;
-            match unsafe { memmap2::Mmap::map(&file) } {
-                Ok(m) => Some(m),
-                Err(e) => {
-                    let _ = tx.send(IndexerMessage::Error {
-                        generation,
-                        message: e.to_string(),
-                    });
-                    return;
-                }
-            }
-        };
-
-        let data = mmap.as_deref().unwrap_or(&[]);
-
-        if !data.is_empty() {
         let mut newline_count: usize = 0;
-            let mut chars_since_check: usize = 0;
+        let mut chunk_offset: u64 = 0;
-            let mut prev_pos: usize = 0;
+        let mut last_byte: Option<u8> = None;
+        let mut bytes_since_check: usize = 0;
 
-            for pos in memchr::memchr_iter(b'\n', data) {
+        loop {
-                chars_since_check += pos - prev_pos;
+            let buf = match buf_reader.fill_buf() {
-                prev_pos = pos;
+                Ok(b) => b,
+                Err(e) => {
+                    send_cancelable(
+                        &tx,
+                        IndexerMessage::Error {
+                            generation,
+                            message: e.to_string(),
+                        },
+                        &cancel_rx,
+                    );
+                    return;
+                }
+            };
+            if buf.is_empty() {
+                break;
+            }
 
-                if chars_since_check >= 1_000_000 {
+            if let Some(&b) = buf.last() {
-                    chars_since_check = 0;
+                last_byte = Some(b);
+            }
+
+            for pos in memchr::memchr_iter(b'\n', buf) {
+                newline_count += 1;
+                if next_line_idx.is_multiple_of(256) {
+                    sampled_offsets.push(chunk_offset + pos as u64 + 1);
+                }
+                next_line_idx += 1;
+            }
+
+            let consumed = buf.len();
+            bytes_since_check += consumed;
+            chunk_offset += consumed as u64;
+            buf_reader.consume(consumed);
+
+            if bytes_since_check >= 1_000_000 {
+                bytes_since_check = 0;
                 if cancel_rx.try_recv().is_ok() {
                     return;
                 }
-                }
+                if target_len > 0 {
-
+                    let percent = (chunk_offset as f64 / target_len as f64) * 100.0;
-                newline_count += 1;
+                    let _ = tx.try_send(IndexerMessage::Progress {
-
-                if newline_count % 256_000 == 0 {
-                    let percent = (pos as f64 / file_size as f64) * 100.0;
-                    let _ = tx.send(IndexerMessage::Progress {
                         generation,
                         percent,
                         lines_scanned: newline_count as u64,
@@ -312,24 +363,84 @@ pub fn spawn_indexer(
             return;
         }
 
-        let line_index = LineIndex::from_bytes(data);
+        let line_index = if chunk_offset == 0 {
+            LineIndex {
+                sampled_offsets: vec![],
+                total_lines: 0,
+                has_trailing_newline: false,
+            }
+        } else {
+            let has_trailing_newline = last_byte == Some(b'\n') && newline_count > 0;
+            let total_lines = if has_trailing_newline && newline_count > 0 {
+                newline_count as u64
+            } else {
+                (1 + newline_count) as u64
+            };
 
-        let _ = IndexCache::save(&path, &line_index);
+            if has_trailing_newline && newline_count > 0 {
+                let trailing_line_idx = newline_count;
+                if trailing_line_idx.is_multiple_of(256) {
+                    sampled_offsets.pop();
+                }
+            }
+
+            LineIndex {
+                sampled_offsets,
+                total_lines,
+                has_trailing_newline,
+            }
+        };
+
+        let mmap = if target_len == 0 {
+            None
+        } else {
+            match std::fs::File::open(&path) {
+                Ok(mmap_file) => match unsafe { memmap2::Mmap::map(&mmap_file) } {
+                    Ok(m) => match mmap_file.metadata() {
+                        Ok(metadata) if metadata.len() >= m.len() as u64 => Some(m),
+                        Ok(_) | Err(_) => None,
+                    },
+                    Err(e) => {
+                        send_cancelable(
+                            &tx,
+                            IndexerMessage::Error {
+                                generation,
+                                message: e.to_string(),
+                            },
+                            &cancel_rx,
+                        );
+                        return;
+                    }
+                },
+                Err(e) => {
+                    send_cancelable(
+                        &tx,
+                        IndexerMessage::Error {
+                            generation,
+                            message: e.to_string(),
+                        },
+                        &cancel_rx,
+                    );
+                    return;
+                }
+            }
+        };
+
+        if let Some(data) = mmap.as_deref() {
+            let _ = IndexCache::save_with_hash(&path, &line_index, data);
+        }
 
         let reader = FileReader::from_parts(path, mmap, line_index);
 
-        let visual_height_index = if terminal_width > 0 {
+        send_cancelable(
-            let visual_heights = compute_visual_heights(&reader, terminal_width, json_format);
+            &tx,
-            Some(VisualHeightIndex::build(&visual_heights).with_params(json_format, terminal_width))
+            IndexerMessage::Complete {
-        } else {
-            None
-        };
-
-        let _ = tx.send(IndexerMessage::Complete {
                 generation,
                 reader,
-            visual_height_index,
+                visual_height_index: None,
-        });
+            },
+            &cancel_rx,
+        );
     });
 
     rx
@@ -355,30 +466,67 @@ pub fn spawn_visual_height_rebuild(
             Err(_) => return,
         };
 
-        let file_size = match file.metadata() {
+        let mmap = match unsafe { memmap2::Mmap::map(&file) } {
-            Ok(m) => m.len(),
+            Ok(m) => match file.metadata() {
+                Ok(meta) if meta.len() >= m.len() as u64 => m,
+                _ => return,
+            },
             Err(_) => return,
         };
+        let data = mmap.as_ref();
+        let total_lines = line_index.line_count();
+        let mut visual_heights: Vec<usize> = Vec::with_capacity(total_lines);
 
-        let mmap = if file_size == 0 {
+        let mut line_start = 0usize;
-            None
+        let mut bytes_since_cancel_check = 0usize;
-        } else {
-            match unsafe { memmap2::Mmap::map(&file) } {
-                Ok(m) => Some(m),
-                Err(_) => return,
-            }
-        };
 
+        loop {
+            if bytes_since_cancel_check >= 1_000_000 {
+                bytes_since_cancel_check = 0;
                 if cancel_rx.try_recv().is_ok() {
                     return;
                 }
+            }
+
+            let newline_rel = memchr::memchr(b'\n', &data[line_start..]);
+            let line_end = match newline_rel {
+                Some(p) => line_start + p,
+                None => data.len(),
+            };
+
+            let mut slice_end = line_end;
+            if slice_end > line_start && data[slice_end - 1] == b'\r' {
+                slice_end -= 1;
+            }
+
+            let line_text = std::str::from_utf8(&data[line_start..slice_end]).unwrap_or("");
+            visual_heights.push(compute_line_visual_height(
+                line_text,
+                terminal_width,
+                json_format,
+            ));
+
+            bytes_since_cancel_check =
+                bytes_since_cancel_check.saturating_add(line_end - line_start + 1);
+
+            match newline_rel {
+                Some(_) => line_start = line_end + 1,
+                None => break,
+            }
+        }
+
+        if visual_heights.len() != total_lines {
+            return;
+        }
 
-        let reader = FileReader::from_parts(path, mmap, line_index);
-        let visual_heights = compute_visual_heights(&reader, terminal_width, json_format);
         let index =
             VisualHeightIndex::build(&visual_heights).with_params(json_format, terminal_width);
 
-        let _ = tx.send(VisualHeightRebuildResult { generation, index });
+        send_cancelable(
+            &tx,
+            VisualHeightRebuildResult { generation, index },
+            &cancel_rx,
+        );
     });
 
     rx
@@ -656,10 +804,10 @@ impl ProgressiveFileReader {
         }
     }
 
-    pub fn update_for_append(&mut self) -> Result<u64> {
+    pub fn update_for_append(&mut self) -> Result<AppendStatus> {
         match &mut self.state {
             ReaderState::Ready { reader, .. } => reader.update_for_append(),
-            _ => Ok(0),
+            _ => Ok(AppendStatus::Unchanged),
         }
     }
 
@@ -675,7 +823,7 @@ impl ProgressiveFileReader {
 
     pub fn start_visual_height_rebuild(&mut self, terminal_width: usize, json_format: bool) {
         if let Some(tx) = self.vh_rebuild_cancel_tx.take() {
-            let _ = tx.send(());
+            let _ = tx.try_send(());
         }
 
         let (cancel_tx, cancel_rx) = crossbeam_channel::bounded(1);
@@ -730,10 +878,10 @@ impl ProgressiveFileReader {
 impl Drop for ProgressiveFileReader {
     fn drop(&mut self) {
         if let Some(tx) = &self.cancel_tx {
-            let _ = tx.send(());
+            let _ = tx.try_send(());
         }
         if let Some(tx) = self.vh_rebuild_cancel_tx.take() {
-            let _ = tx.send(());
+            let _ = tx.try_send(());
         }
     }
 }
@@ -991,12 +1139,10 @@ mod tests {
                 assert_eq!(reader.get_line(0), Some("line1"));
                 assert_eq!(reader.get_line(1), Some("line2"));
                 assert_eq!(reader.get_line(2), Some("line3"));
-                assert!(visual_height_index.is_some());
+                assert!(
-                let idx = visual_height_index.unwrap();
+                    visual_height_index.is_none(),
-                assert_eq!(idx.total_visual_rows(), 3);
+                    "spawn_indexer no longer builds VHI inline; UI triggers rebuild post-layout"
-                assert_eq!(idx.visual_row_to_logical_row(0), 0);
+                );
-                assert_eq!(idx.visual_row_to_logical_row(1), 1);
-                assert_eq!(idx.visual_row_to_logical_row(2), 2);
             }
             other => panic!("expected Complete, got {:?}", other),
         }
@@ -1193,20 +1339,25 @@ mod tests {
         let (_cancel_tx, cancel_rx) = crossbeam_channel::bounded(1);
 
         let rx = spawn_indexer(temp.path().to_path_buf(), 1, 80, false, cancel_rx);
-        let msg = rx.recv_timeout(std::time::Duration::from_secs(10)).unwrap();
+        loop {
-
+            match rx.recv_timeout(std::time::Duration::from_secs(10)).unwrap() {
-        match msg {
+                IndexerMessage::Progress { .. } => continue,
                 IndexerMessage::Complete {
+                    reader,
                     visual_height_index,
                     ..
                 } => {
-                let idx = visual_height_index.expect("should have visual height index");
+                    assert_eq!(reader.line_count(), 2);
-                assert_eq!(idx.visual_height_of_line(0), 1);
+                    assert!(
-                assert_eq!(idx.visual_height_of_line(1), 1);
+                        visual_height_index.is_none(),
+                        "spawn_indexer no longer builds VHI inline; UI triggers rebuild post-layout"
+                    );
+                    break;
                 }
                 other => panic!("expected Complete, got {:?}", other),
             }
         }
+    }
 
     #[test]
     fn test_visual_height_index_is_valid_for() {
@@ -1258,6 +1409,247 @@ mod tests {
         idx.extend_from_heights(&[1, 2, 3]);
 
         assert_eq!(idx.total_visual_rows(), 6);
+    }
+
+    #[test]
+    fn test_replace_last_line_height_increase() {
+        let heights = [2, 3];
+        let mut idx = VisualHeightIndex::build(&heights);
+        assert_eq!(idx.total_visual_rows(), 5);
+        assert_eq!(idx.visual_height_of_line(1), 3);
+
+        idx.replace_last_line_height(7);
+
+        assert_eq!(idx.visual_height_of_line(0), 2);
+        assert_eq!(idx.visual_height_of_line(1), 7);
+        assert_eq!(idx.total_visual_rows(), 9);
+        assert_eq!(idx.cursor_to_first_visual_row(0), 0);
+        assert_eq!(idx.cursor_to_first_visual_row(1), 2);
+    }
+
+    #[test]
+    fn test_replace_last_line_height_decrease() {
+        let heights = [2, 5];
+        let mut idx = VisualHeightIndex::build(&heights);
+
+        idx.replace_last_line_height(1);
+
+        assert_eq!(idx.visual_height_of_line(0), 2);
+        assert_eq!(idx.visual_height_of_line(1), 1);
+        assert_eq!(idx.total_visual_rows(), 3);
+    }
+
+    #[test]
+    fn test_replace_last_line_height_same_is_noop() {
+        let heights = [2, 3];
+        let mut idx = VisualHeightIndex::build(&heights);
+        let total_before = idx.total_visual_rows();
+
+        idx.replace_last_line_height(3);
+
+        assert_eq!(idx.total_visual_rows(), total_before);
+    }
+
+    #[test]
+    fn test_replace_last_line_height_then_extend() {
+        let heights = [2, 1];
+        let mut idx = VisualHeightIndex::build(&heights);
+        assert_eq!(idx.total_visual_rows(), 3);
+
+        idx.replace_last_line_height(4);
+        idx.extend_from_heights(&[3]);
+
         assert_eq!(idx.line_count(), 3);
+        assert_eq!(idx.visual_height_of_line(0), 2);
+        assert_eq!(idx.visual_height_of_line(1), 4);
+        assert_eq!(idx.visual_height_of_line(2), 3);
+        assert_eq!(idx.total_visual_rows(), 9);
+        assert_eq!(idx.cursor_to_first_visual_row(0), 0);
+        assert_eq!(idx.cursor_to_first_visual_row(1), 2);
+        assert_eq!(idx.cursor_to_first_visual_row(2), 6);
+    }
+
+    #[test]
+    fn test_replace_last_line_height_single_line() {
+        let heights = [5];
+        let mut idx = VisualHeightIndex::build(&heights);
+
+        idx.replace_last_line_height(2);
+
+        assert_eq!(idx.visual_height_of_line(0), 2);
+        assert_eq!(idx.total_visual_rows(), 2);
+    }
+
+    #[test]
+    fn test_replace_last_line_height_empty_index() {
+        let heights: [usize; 0] = [];
+        let mut idx = VisualHeightIndex::build(&heights);
+
+        idx.replace_last_line_height(5);
+
+        assert_eq!(idx.total_visual_rows(), 0);
+    }
+
+    #[test]
+    fn test_spawn_indexer_file_truncated_during_scan() {
+        let mut content = Vec::new();
+        for i in 0..100_000 {
+            writeln!(content, "line number {:08}", i).unwrap();
+        }
+        let f = create_temp_file(&content);
+        let (_cancel_tx, cancel_rx) = crossbeam_channel::bounded(1);
+
+        let rx = spawn_indexer(f.path().to_path_buf(), 1, 80, false, cancel_rx);
+
+        {
+            use std::io::Write;
+            let _ = std::fs::OpenOptions::new()
+                .write(true)
+                .truncate(true)
+                .open(f.path())
+                .unwrap();
+        }
+
+        let result = rx.recv_timeout(std::time::Duration::from_secs(10));
+        match result {
+            Ok(IndexerMessage::Complete { reader, .. }) => {
+                assert!(reader.line_count() <= 100_000);
+            }
+            Ok(IndexerMessage::Error { .. }) => {}
+            Err(crossbeam_channel::RecvTimeoutError::Disconnected) => {}
+            Err(crossbeam_channel::RecvTimeoutError::Timeout) => {}
+            other => panic!("unexpected: {:?}", other),
+        }
+    }
+
+    #[test]
+    fn test_spawn_visual_height_rebuild_line_count_mismatch_discards() {
+        let content = b"line0\nline1\nline2\n";
+        let f = create_temp_file(content);
+        let data = std::fs::read(f.path()).unwrap();
+        let index = LineIndex::from_bytes(&data);
+        IndexCache::save(f.path(), &index).unwrap();
+
+        {
+            use std::io::Write;
+            let mut file = std::fs::OpenOptions::new()
+                .write(true)
+                .truncate(true)
+                .open(f.path())
+                .unwrap();
+            file.write_all(b"only_one_line\n").unwrap();
+        }
+
+        let (_cancel_tx, cancel_rx) = crossbeam_channel::bounded(1);
+        let rx = spawn_visual_height_rebuild(f.path().to_path_buf(), 1, 80, false, cancel_rx);
+
+        let result = rx.recv_timeout(std::time::Duration::from_secs(5));
+        match result {
+            Err(crossbeam_channel::RecvTimeoutError::Disconnected) => {}
+            Err(crossbeam_channel::RecvTimeoutError::Timeout) => {}
+            Ok(_) => panic!("should have been discarded due to line count mismatch"),
+        }
+    }
+
+    #[test]
+    fn test_send_cancelable_delivers_on_empty_channel() {
+        let (tx, rx) = crossbeam_channel::bounded(2);
+        let (_cancel_tx, cancel_rx) = crossbeam_channel::bounded(1);
+
+        send_cancelable(&tx, 42, &cancel_rx);
+
+        assert_eq!(rx.try_recv(), Ok(42));
+    }
+
+    #[test]
+    fn test_send_cancelable_aborts_on_cancel() {
+        let (tx, rx) = crossbeam_channel::bounded(1);
+        let (cancel_tx, cancel_rx) = crossbeam_channel::bounded(1);
+
+        tx.send("filler").unwrap();
+
+        let handle = std::thread::spawn(move || {
+            send_cancelable(&tx, "important", &cancel_rx);
+        });
+
+        cancel_tx.send(()).unwrap();
+        handle.join().unwrap();
+
+        assert_eq!(rx.try_recv(), Ok("filler"));
+    }
+
+    #[test]
+    fn test_send_cancelable_drains_when_room_available() {
+        let (tx, rx) = crossbeam_channel::bounded(1);
+        let (_cancel_tx, cancel_rx) = crossbeam_channel::bounded(1);
+
+        tx.send("first").unwrap();
+
+        let rx_clone = rx.clone();
+        let handle = std::thread::spawn(move || {
+            std::thread::sleep(std::time::Duration::from_millis(50));
+            let _ = rx_clone.try_recv();
+        });
+
+        send_cancelable(&tx, "second", &cancel_rx);
+        handle.join().unwrap();
+
+        assert_eq!(rx.try_recv(), Ok("second"));
+    }
+
+    #[test]
+    fn test_progress_try_send_does_not_block_full_channel() {
+        let mut content = Vec::new();
+        for i in 0..50_000 {
+            writeln!(content, "line number {:08}", i).unwrap();
+        }
+        let f = create_temp_file(&content);
+        let (_cancel_tx, cancel_rx) = crossbeam_channel::bounded(1);
+
+        let rx = spawn_indexer(f.path().to_path_buf(), 1, 80, false, cancel_rx);
+
+        let mut got_complete = false;
+        let timeout = std::time::Duration::from_secs(15);
+        let start = std::time::Instant::now();
+        while start.elapsed() < timeout {
+            match rx.recv_timeout(std::time::Duration::from_secs(1)) {
+                Ok(IndexerMessage::Progress { .. }) => {}
+                Ok(IndexerMessage::Complete { .. }) => {
+                    got_complete = true;
+                    break;
+                }
+                Ok(IndexerMessage::Error { message, .. }) => {
+                    panic!("unexpected error: {}", message);
+                }
+                Err(e) => panic!("recv error: {:?}", e),
+            }
+        }
+        assert!(
+            got_complete,
+            "indexer should complete even when Progress fills channel"
+        );
+    }
+
+    #[test]
+    fn test_indexer_cancel_with_full_channel() {
+        let mut content = Vec::new();
+        for i in 0..500_000 {
+            writeln!(content, "line number {:08}", i).unwrap();
+        }
+        let f = create_temp_file(&content);
+        let (cancel_tx, cancel_rx) = crossbeam_channel::bounded(1);
+
+        let rx = spawn_indexer(f.path().to_path_buf(), 1, 80, false, cancel_rx);
+
+        cancel_tx.send(()).unwrap();
+
+        let result = rx.recv_timeout(std::time::Duration::from_secs(5));
+        match result {
+            Err(crossbeam_channel::RecvTimeoutError::Timeout)
+            | Err(crossbeam_channel::RecvTimeoutError::Disconnected) => {}
+            Ok(IndexerMessage::Complete { .. }) => {}
+            Ok(IndexerMessage::Error { .. }) => {}
+            Ok(IndexerMessage::Progress { .. }) => {}
+        }
     }
 }

crates/core/src/io/read_cache.rs

@@ -56,9 +56,15 @@ impl LruReadCache {
     /// on a hit, or fills a cache slot on a miss. Cross-block reads go through
     /// the spill buffer and are not cached.
     pub fn get(&mut self, file: &File, offset: u64, len: usize) -> io::Result<&[u8]> {
+        if len == 0 {
+            return Ok(&[]);
+        }
+
         let aligned_key = offset & !(BLOCK_ALIGN as u64 - 1);
-        let request_end = offset.saturating_add(len as u64);
+        let request_end = offset.checked_add(len as u64).ok_or_else(|| {
-        let block_end = aligned_key + BLOCK_ALIGN as u64;
+            io::Error::new(io::ErrorKind::InvalidInput, "read range overflows u64")
+        })?;
+        let block_end = aligned_key.saturating_add(BLOCK_ALIGN as u64);
 
         if request_end > block_end {
             self.spill_buf.resize(len, 0);
@@ -74,7 +80,8 @@ impl LruReadCache {
         }
 
         let hit_idx = self.slots.iter().position(|slot| {
-            slot.block_offset == aligned_key && request_end <= slot.block_offset + slot.len as u64
+            let slot_end = slot.block_offset.saturating_add(slot.len as u64);
+            slot.len > 0 && slot.block_offset == aligned_key && request_end <= slot_end
         });
 
         if let Some(idx) = hit_idx {
@@ -96,8 +103,8 @@ impl LruReadCache {
         let slot = &mut self.slots[evict_idx];
         let bytes_read = file.read_at(&mut slot.buf, aligned_key)?;
 
-        // Note: get(file, 0, 0) on an empty file now returns Err (old code returned Ok(&[])).
+        // Non-empty reads that return 0 are EOF. Zero-length reads are handled above
-        // No callers pass len == 0, so this is a safe semantic change.
+        // as a successful no-op.
         if bytes_read == 0 {
             return Err(io::Error::new(io::ErrorKind::UnexpectedEof, "read 0 bytes"));
         }
@@ -107,7 +114,8 @@ impl LruReadCache {
         slot.last_access = self.tick;
         self.tick += 1;
 
-        if request_end > aligned_key + bytes_read as u64 {
+        let bytes_end = aligned_key.saturating_add(bytes_read as u64);
+        if request_end > bytes_end {
             return Err(io::Error::new(io::ErrorKind::UnexpectedEof, "short read"));
         }
 
@@ -118,7 +126,9 @@ impl LruReadCache {
     /// Invalidate all cache slots and the spill buffer.
     pub fn clear(&mut self) {
         for slot in &mut self.slots {
+            slot.block_offset = 0;
             slot.len = 0;
+            slot.last_access = 0;
         }
         self.spill_len = 0;
     }
@@ -314,9 +324,11 @@ mod tests {
 
         cache.clear();
 
-        // All slots should have len == 0.
+        // All slots should be fully reset.
         for slot in &cache.slots {
+            assert_eq!(slot.block_offset, 0);
             assert_eq!(slot.len, 0);
+            assert_eq!(slot.last_access, 0);
         }
         assert_eq!(cache.spill_len, 0);
 
@@ -432,4 +444,52 @@ mod tests {
         assert_eq!(&line2[..4090], &[b'B'; 4090]);
         assert_eq!(line2[4090], b'\n');
     }
+
+    #[test]
+    fn zero_len_read_is_noop_on_fresh_cache() {
+        let f = make_file(b"");
+        let file = File::open(f.path()).unwrap();
+        let mut cache = ReadCache::new();
+
+        let result = cache.get(&file, 0, 0).unwrap();
+        assert!(result.is_empty());
+        assert_eq!(cache.tick, 0);
+        assert!(cache.slots.iter().all(|s| s.len == 0));
+    }
+
+    #[test]
+    fn zero_len_read_is_noop_on_populated_cache() {
+        let f = make_file(b"abc");
+        let file = File::open(f.path()).unwrap();
+        let mut cache = ReadCache::new();
+
+        cache.get(&file, 0, 1).unwrap();
+        let tick_before = cache.tick;
+
+        let result = cache.get(&file, 0, 0).unwrap();
+        assert!(result.is_empty());
+        assert_eq!(cache.tick, tick_before);
+    }
+
+    #[test]
+    fn zero_len_read_at_max_offset_is_ok() {
+        let f = make_file(b"");
+        let file = File::open(f.path()).unwrap();
+        let mut cache = ReadCache::new();
+
+        let result = cache.get(&file, u64::MAX, 0).unwrap();
+        assert!(result.is_empty());
+        assert_eq!(cache.tick, 0);
+        assert!(cache.slots.iter().all(|s| s.len == 0));
+    }
+
+    #[test]
+    fn nonzero_read_range_overflow_returns_invalid_input() {
+        let f = make_file(b"abc");
+        let file = File::open(f.path()).unwrap();
+        let mut cache = ReadCache::new();
+
+        let err = cache.get(&file, u64::MAX, 1).unwrap_err();
+        assert_eq!(err.kind(), io::ErrorKind::InvalidInput);
+    }
 }

crates/core/src/io/wrap.rs

@@ -1,10 +1,21 @@
+use std::borrow::Cow;
+
 /// Maximum input length for wrap/format operations (10 MB).
-/// Lines exceeding this are returned as-is to avoid pathological cases.
+/// Callers should check against this constant before invoking `wrap_line_chars`
+/// to avoid pathological cases on oversized lines.
 pub const MAX_WRAP_INPUT_LEN: usize = 10 * 1024 * 1024;
 
-/// Split a line into chunks of exactly `width` characters (display columns).
+/// Column spacing for tab stop alignment.
+const TAB_WIDTH: usize = 4;
+
+/// Split a line into chunks of exactly `width` display columns.
 /// For a log viewer, we want character-level wrapping, not word-level.
+/// Uses `unicode-width` for correct CJK/emoji/zero-width handling.
+/// Tab characters expand to the next tab-stop boundary and split across
+/// rows when the expansion exceeds the remaining width.
 pub fn wrap_line_chars(line: &str, width: usize) -> Vec<String> {
+    use unicode_width::UnicodeWidthChar;
+
     if width == 0 {
         return vec![String::new()];
     }
@@ -15,23 +26,42 @@ pub fn wrap_line_chars(line: &str, width: usize) -> Vec<String> {
     let mut row = String::new();
     let mut col = 0;
     for ch in line.chars() {
-        let w = if ch == '\t' { 4 } else { 1 };
+        if ch == '\t' {
+            let tab_stop = TAB_WIDTH - (col % TAB_WIDTH);
+            let mut remaining = tab_stop;
+            while remaining > 0 {
+                let avail = width.saturating_sub(col);
+                let fill = remaining.min(avail);
+                for _ in 0..fill {
+                    row.push(' ');
+                }
+                col += fill;
+                remaining -= fill;
+                if col >= width {
+                    result.push(std::mem::take(&mut row));
+                    col = 0;
+                }
+            }
+        } else {
+            let w = if ch.is_control() {
+                // Control characters (except tab): width 0, still pushed to preserve content.
+                // Visible rendering is the caller's responsibility.
+                0
+            } else {
+                ch.width().unwrap_or(0)
+            };
             if col + w > width && !row.is_empty() {
                 result.push(std::mem::take(&mut row));
                 col = 0;
             }
-        if ch == '\t' {
-            row.push_str("    ");
-            col += 4;
-        } else {
             row.push(ch);
             col += w;
-        }
             if col >= width {
                 result.push(std::mem::take(&mut row));
                 col = 0;
             }
         }
+    }
     if !row.is_empty() {
         result.push(row);
     }
@@ -41,21 +71,81 @@ pub fn wrap_line_chars(line: &str, width: usize) -> Vec<String> {
     result
 }
 
-/// Format a line as pretty-printed JSON if it's a JSON Object.
+/// Count wrapped rows for a line without allocating the wrapped strings.
-/// Returns the original line unchanged for non-JSON or non-Object content.
+/// MUST produce the same count as `wrap_line_chars(line, width).len()`.
-pub fn format_json_line(line: &str) -> String {
+pub fn wrap_line_count(line: &str, width: usize) -> usize {
-    if line.trim().is_empty() {
+    use unicode_width::UnicodeWidthChar;
-        return String::new();
+
+    if width == 0 || line.is_empty() {
+        return 1;
     }
-    // Quick pre-check: only try parsing if it starts with '{'
+    let mut count = 0usize;
-    if !line.trim_start().starts_with('{') {
+    let mut col = 0usize;
-        return line.to_string();
+    let mut row_has_content = false;
+
+    for ch in line.chars() {
+        if ch == '\t' {
+            let tab_stop = TAB_WIDTH - (col % TAB_WIDTH);
+            let mut remaining = tab_stop;
+            while remaining > 0 {
+                let avail = width.saturating_sub(col);
+                let fill = remaining.min(avail);
+                col += fill;
+                remaining -= fill;
+                if col >= width {
+                    count += 1;
+                    col = 0;
+                    row_has_content = false;
+                } else {
+                    row_has_content = true;
+                }
+            }
+        } else {
+            let w = if ch.is_control() {
+                0
+            } else {
+                ch.width().unwrap_or(0)
+            };
+            if col + w > width && row_has_content {
+                count += 1;
+                col = 0;
+            }
+            col += w;
+            if col >= width {
+                count += 1;
+                col = 0;
+                row_has_content = false;
+            } else {
+                row_has_content = true;
+            }
+        }
+    }
+    if row_has_content {
+        count += 1;
+    }
+    if count == 0 {
+        count = 1;
+    }
+    count
+}
+
+/// Format a line as pretty-printed JSON if it's a JSON Object.
+/// Returns the original line borrowed for non-JSON or non-Object content,
+/// only allocating when pretty-printing actually happens.
+pub fn format_json_line(line: &str) -> Cow<'_, str> {
+    let trimmed = line.trim();
+    if trimmed.is_empty() {
+        return Cow::Borrowed("");
+    }
+    if !trimmed.starts_with('{') {
+        return Cow::Borrowed(line);
     }
     match serde_json::from_str::<serde_json::Value>(line) {
-        Ok(value) if value.is_object() => {
+        Ok(value) if value.is_object() => match serde_json::to_string_pretty(&value) {
-            serde_json::to_string_pretty(&value).unwrap_or_else(|_| line.to_string())
+            Ok(s) => Cow::Owned(s),
-        }
+            Err(_) => Cow::Borrowed(line),
-        _ => line.to_string(),
+        },
+        _ => Cow::Borrowed(line),
     }
 }
 
@@ -96,7 +186,7 @@ mod tests {
     #[test]
     fn test_wrap_with_tab() {
         let result = wrap_line_chars("a\tb", 4);
-        assert_eq!(result, vec!["a", "    ", "b"]);
+        assert_eq!(result, vec!["a   ", "b"]);
     }
 
     #[test]
@@ -132,4 +222,195 @@ mod tests {
     fn test_max_wrap_input_len_constant() {
         assert_eq!(MAX_WRAP_INPUT_LEN, 10 * 1024 * 1024);
     }
+
+    #[test]
+    fn test_wrap_cjk_chars() {
+        let result = wrap_line_chars("你好", 3);
+        assert_eq!(result, vec!["你", "好"]);
+    }
+
+    #[test]
+    fn test_wrap_cjk_ascii_mixed() {
+        let result = wrap_line_chars("a你好", 4);
+        assert_eq!(result, vec!["a你", "好"]);
+    }
+
+    #[test]
+    fn test_wrap_zero_width_char() {
+        let result = wrap_line_chars("a\u{200B}b", 2);
+        assert_eq!(result, vec!["a\u{200B}b"]);
+    }
+
+    #[test]
+    fn test_wrap_emoji() {
+        let result = wrap_line_chars("😀a", 3);
+        assert_eq!(result, vec!["😀a"]);
+    }
+
+    #[test]
+    fn test_wrap_emoji_exact_wrap() {
+        let result = wrap_line_chars("😀a", 2);
+        assert_eq!(result, vec!["😀", "a"]);
+    }
+
+    #[test]
+    fn test_wrap_combining_mark() {
+        // Scalar-width wrapping: combining mark (width 0) stays with next base char,
+        // not the preceding one, because the base char already triggered a flush.
+        let result = wrap_line_chars("a\u{0301}b", 1);
+        assert_eq!(result, vec!["a", "\u{0301}b"]);
+    }
+
+    #[test]
+    fn test_wrap_cjk_width_one() {
+        let result = wrap_line_chars("你好", 1);
+        assert_eq!(result, vec!["你", "好"]);
+    }
+
+    #[test]
+    fn test_tab_narrow_width() {
+        let result = wrap_line_chars("\t", 2);
+        assert_eq!(result, vec!["  ", "  "]);
+        let result = wrap_line_chars("\t", 1);
+        assert_eq!(result, vec![" ", " ", " ", " "]);
+    }
+
+    #[test]
+    fn test_tab_stop_alignment() {
+        assert_eq!(wrap_line_chars("a\tb", 8), vec!["a   b"]);
+        assert_eq!(wrap_line_chars("ab\t", 4), vec!["ab  "]);
+        assert_eq!(wrap_line_chars("abc\tb", 8), vec!["abc b"]);
+    }
+
+    #[test]
+    fn test_tab_at_line_boundary() {
+        let result = wrap_line_chars("a\tb", 4);
+        assert_eq!(result, vec!["a   ", "b"]);
+    }
+
+    #[test]
+    fn test_tab_regression_ab_tab() {
+        let result = wrap_line_chars("ab\t", 4);
+        assert_eq!(result, vec!["ab  "]);
+    }
+
+    #[test]
+    fn test_wrap_count_matches_chars_empty() {
+        assert_eq!(wrap_line_count("", 80), wrap_line_chars("", 80).len());
+    }
+
+    #[test]
+    fn test_wrap_count_matches_chars_zero_width() {
+        assert_eq!(
+            wrap_line_count("hello", 0),
+            wrap_line_chars("hello", 0).len()
+        );
+    }
+
+    #[test]
+    fn test_wrap_count_matches_chars_ascii_short() {
+        for width in 1..20 {
+            assert_eq!(
+                wrap_line_count("hello world", width),
+                wrap_line_chars("hello world", width).len(),
+                "width={width}"
+            );
+        }
+    }
+
+    #[test]
+    fn test_wrap_count_matches_chars_cjk() {
+        for width in 1..10 {
+            assert_eq!(
+                wrap_line_count("你好世界", width),
+                wrap_line_chars("你好世界", width).len(),
+                "width={width}"
+            );
+        }
+    }
+
+    #[test]
+    fn test_wrap_count_matches_chars_emoji() {
+        for width in 1..6 {
+            assert_eq!(
+                wrap_line_count("😀👨‍👩‍👧", width),
+                wrap_line_chars("😀👨‍👩‍👧", width).len(),
+                "width={width}"
+            );
+        }
+    }
+
+    #[test]
+    fn test_wrap_count_matches_chars_combining_mark() {
+        for width in 1..6 {
+            assert_eq!(
+                wrap_line_count("a\u{0301}b\u{0301}c", width),
+                wrap_line_chars("a\u{0301}b\u{0301}c", width).len(),
+                "width={width}"
+            );
+        }
+    }
+
+    #[test]
+    fn test_wrap_count_matches_chars_tab_variants() {
+        let inputs = ["\t", "a\tb", "ab\t", "\t\t", "abc\tb", "a\t\t\tb"];
+        for input in inputs {
+            for width in 1..9 {
+                assert_eq!(
+                    wrap_line_count(input, width),
+                    wrap_line_chars(input, width).len(),
+                    "input={input:?} width={width}"
+                );
+            }
+        }
+    }
+
+    #[test]
+    fn test_wrap_count_matches_chars_mixed() {
+        let input = "log: 你好 😀\tvalue=true [ERROR]";
+        for width in 1..40 {
+            assert_eq!(
+                wrap_line_count(input, width),
+                wrap_line_chars(input, width).len(),
+                "width={width}"
+            );
+        }
+    }
+
+    #[test]
+    fn test_wrap_count_matches_chars_zero_width_chars() {
+        let input = "a\u{200B}b\u{200C}c\u{200D}d";
+        for width in 1..8 {
+            assert_eq!(
+                wrap_line_count(input, width),
+                wrap_line_chars(input, width).len(),
+                "width={width}"
+            );
+        }
+    }
+
+    #[test]
+    fn test_wrap_count_matches_chars_control_chars() {
+        let input = "a\x07b\x1Fc";
+        for width in 1..8 {
+            assert_eq!(
+                wrap_line_count(input, width),
+                wrap_line_chars(input, width).len(),
+                "width={width}"
+            );
+        }
+    }
+
+    #[test]
+    fn test_wrap_count_matches_chars_long_log_line() {
+        let input = "2025-01-14T10:23:45.123Z [INFO] [auth] request handled status=200 \
+                     latency=0.234s endpoint=/api/v1/users user=user_42 request_id=req_abc123def456";
+        for width in [10, 20, 40, 60, 80, 100, 120] {
+            assert_eq!(
+                wrap_line_count(input, width),
+                wrap_line_chars(input, width).len(),
+                "width={width}"
+            );
+        }
+    }
 }

crates/core/src/parser/json.rs

@@ -16,7 +16,14 @@
 // 类似于 Python 的 dict、JavaScript 的 Object/Map、Java 的 HashMap。
 // 它存储键值对（key-value pairs），可以通过键快速查找对应的值。
 // 这里用 HashMap<String, Value> 来存储 JSON 中除 timestamp/level 之外的其他字段。
-use std::collections::HashMap;
+use std::collections::{HashMap, HashSet};
+
+// serde::de 中的 Visitor / MapAccess 允许我们自定义 JSON 对象的反序列化过程。
+// 默认的 serde_json::from_str::<HashMap<_, _>>() 遇到重复键时会采用"后者覆盖前者"（last-wins），
+// 前面的值被静默丢弃。这里我们通过自定义 Visitor 在反序列化过程中逐个观察 key-value 对，
+// 在保持 last-wins 行为的同时，将重复 key 的所有值记录到 DuplicateKey 中。
+use serde::Deserializer;
+use serde::de::{MapAccess, Visitor};
 
 // serde_json::Value — 来自 serde_json 库（Rust 中最流行的 JSON 处理库）。
 // Value 是一个枚举类型，可以表示任意 JSON 值：
@@ -32,7 +39,20 @@ use serde_json::Value;
 // ─── 引入项目内部类型 ──────────────────────────────────────────────────────
 // crate 表示"当前项目（crate）"。
 // types 模块中定义了 LogEntry（一条日志记录）和 LogLevel（日志级别，如 INFO/ERROR）。
-use crate::types::{LogEntry, LogLevel};
+use crate::types::{DuplicateKey, LogEntry, LogLevel};
+
+// ─── strip_bom 辅助函数 ──────────────────────────────────────────────────
+// 剥离行首的 UTF-8 BOM（Byte Order Mark, U+FEFF）。
+//
+// Windows 环境和某些导出工具生成的文件会在行首插入 BOM，
+// 而 serde_json 不接受 BOM 前缀的 JSON 文本（会报 "expected value" 错误）。
+// 只剥离一个前导 BOM，不处理多个 BOM 或行内 BOM（那些是畸形输入）。
+//
+// 参数: line: &str — 输入字符串切片。
+// 返回: &str — 去掉 BOM 后的字符串切片（借用原始字符串，零分配）。
+fn strip_bom(line: &str) -> &str {
+    line.strip_prefix('\u{FEFF}').unwrap_or(line)
+}
 
 // ─── detect_json_log 函数 ──────────────────────────────────────────────────
 // 检测一行文本是否是一个 JSON 对象。
@@ -63,7 +83,92 @@ pub fn detect_json_log(line: &str) -> bool {
     //    则匹配成功。_ 是通配符，表示"不关心对象里面的具体内容"。
     //
     // 如果匹配到 Ok(Value::Object(_)) 返回 true，否则返回 false。
-    matches!(serde_json::from_str::<Value>(line), Ok(Value::Object(_)))
+    matches!(
+        serde_json::from_str::<Value>(strip_bom(line)),
+        Ok(Value::Object(_))
+    )
+}
+
+// ─── DuplicateKeyVisitor ──────────────────────────────────────────────────
+// 自定义 serde Visitor，在反序列化 JSON 对象时检测重复 key。
+//
+// 工作原理：
+//   serde 的 MapAccess trait 允许我们逐个遍历 JSON 对象的 key-value 对。
+//   每读到一个 (key, value)，我们：
+//     1. 检查这个 key 是否已经见过（通过 HashSet）
+//     2. 如果是重复 key，记录到 Vec<DuplicateKey> 中（包含所有出现过的值）
+//     3. 将 key-value 插入 Map（last-wins，与 serde_json 默认行为一致）
+//
+// 这样既保持了兼容性（last-wins），又不丢失信息（所有值都记录在 DuplicateKey 中）。
+struct DuplicateKeyVisitor;
+
+impl<'de> Visitor<'de> for DuplicateKeyVisitor {
+    // 返回类型：(serde_json::Map, 重复 key 列表)
+    type Value = (serde_json::Map<String, Value>, Vec<DuplicateKey>);
+
+    fn expecting(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
+        f.write_str("a JSON object")
+    }
+
+    fn visit_map<A>(self, mut access: A) -> Result<Self::Value, A::Error>
+    where
+        A: MapAccess<'de>,
+    {
+        let mut map = serde_json::Map::new();
+        let mut seen = HashSet::new();
+        let mut duplicates: Vec<DuplicateKey> = Vec::new();
+
+        while let Some((key, value)) = access.next_entry::<String, Value>()? {
+            if !seen.insert(key.clone()) {
+                // 重复 key：将之前 map 中的值和当前值都记录下来
+                if let Some(existing) = duplicates.iter_mut().find(|d| d.key == key) {
+                    // 同一个 key 第三次及以上出现：追加当前值
+                    existing.values.push(value.clone());
+                } else {
+                    // 同一个 key 第二次出现：记录第一次的值 + 当前值
+                    let prev_value = map.get(&key).cloned().unwrap_or(Value::Null);
+                    duplicates.push(DuplicateKey {
+                        key: key.clone(),
+                        values: vec![prev_value, value.clone()],
+                    });
+                }
+            }
+            // last-wins：后出现的值覆盖前面的值，与 serde_json 默认行为一致
+            map.insert(key, value);
+        }
+
+        Ok((map, duplicates))
+    }
+}
+
+/// 使用自定义 Visitor 解析 JSON 对象，同时检测重复 key。
+///
+/// 返回 (serde_json::Map, Vec<DuplicateKey>)：
+///   - Map 中存储所有 key-value（重复 key 取 last-wins）
+///   - Vec 中记录所有重复 key 及其全部值
+fn parse_json_object_with_duplicates(
+    json: &str,
+) -> Option<(serde_json::Map<String, Value>, Vec<DuplicateKey>)> {
+    let mut deserializer = serde_json::Deserializer::from_str(json);
+    deserializer.deserialize_map(DuplicateKeyVisitor).ok()
+}
+
+// ─── take_string_field_from_map 辅助函数 ──────────────────────────────────
+// 从 serde_json::Map 中安全提取字符串字段。
+// 功能与原 take_string_field 相同，但操作 serde_json::Map 而非 HashMap。
+fn take_string_field_from_map(
+    obj: &mut serde_json::Map<String, Value>,
+    keys: &[&str],
+) -> Option<String> {
+    for key in keys {
+        if obj.get(*key).is_some_and(Value::is_string) {
+            let Some(Value::String(v)) = obj.remove(*key) else {
+                unreachable!("value was checked as string");
+            };
+            return Some(v);
+        }
+    }
+    None
 }
 
 // ─── parse_line 函数 ──────────────────────────────────────────────────────
@@ -73,107 +178,35 @@ pub fn detect_json_log(line: &str) -> bool {
 // 返回: Option<LogEntry> — 解析成功返回 Some(LogEntry)，失败或不合法返回 None。
 //   Option 是 Rust 的可选类型：Some(值) 表示有值，None 表示没有值。
 pub fn parse_line(line: &str) -> Option<LogEntry> {
-    // ─── 跳过空行 ──────────────────────────────────────────────────────────
+    let line = strip_bom(line);
-    // line.trim() 去除首尾空白字符（空格、制表符、换行符等）。
+
-    // .is_empty() 检查是否为空字符串。
-    // 如果去除空白后是空的，说明是空行，不需要解析，直接返回 None。
     if line.trim().is_empty() {
         return None;
     }
 
-    // ─── 解析 JSON 为 HashMap ──────────────────────────────────────────────
+    // ─── 使用自定义 Visitor 解析 JSON ──────────────────────────────────
-    // serde_json::from_str(line) 尝试将字符串解析为 JSON。
+    // 通过 DuplicateKeyVisitor 反序列化，在保持 last-wins 的同时检测重复 key。
-    // 由于我们声明了 HashMap<String, Value> 类型，Rust 会自动将 JSON 对象
+    // 返回的 (serde_json::Map, Vec<DuplicateKey>) 中：
-    // 转换为 HashMap，其中每个键是 String，每个值是 serde_json::Value。
+    //   - Map 包含所有 key-value（重复 key 取最后一个值）
-    //
+    //   - Vec 记录了所有重复 key 及其出现过的全部值
-    // .ok() 将 Result 转换为 Option：
+    let (mut obj, duplicate_keys) = parse_json_object_with_duplicates(line)?;
-    //   Ok(值) → Some(值)
-    //   Err(_) → None
-    //
-    // 末尾的 ? 是"问号操作符"（try operator），在这里的作用是：
-    // 如果 .ok() 返回 None（即 JSON 解析失败），则整个函数直接返回 None。
-    // 如果返回 Some(hashmap)，则将 hashmap 取出并绑定到 fields 变量。
-    //
-    // let mut 表示这是一个"可变变量"（mutable variable），
-    // 后续代码会修改这个 HashMap（从中删除已识别的字段）。
-    let mut fields: HashMap<String, Value> = serde_json::from_str(line).ok()?;
 
-    // ─── 保存原始行内容 ──────────────────────────────────────────────────
-    // line.to_string() 将 &str（字符串切片引用）转换为 String（拥有所有权的字符串）。
-    // 保存原始行是为了在 UI 中显示未经修改的原始日志内容。
     let raw_line = line.to_string();
-
+    let timestamp =
-    // ─── 提取时间戳字段 ──────────────────────────────────────────────────
+        take_string_field_from_map(&mut obj, &["timestamp", "time", "ts", "@timestamp"]);
-    // 这段代码尝试从 JSON 中提取时间戳，逻辑如下：
+    let level = take_string_field_from_map(&mut obj, &["level", "lvl", "severity"])
-    //
-    // 1. ["timestamp", "time", "ts", "@timestamp"] — 候选键名数组。
-    //    不同日志系统使用不同的时间戳字段名，这里列出常见的几种。
-    //
-    // 2. .iter() — 创建数组的迭代器，可以逐个遍历元素。
-    //
-    // 3. .find_map(|key| fields.remove(*key)) — 对每个候选键名：
-    //    - fields.remove(*key): 尝试从 HashMap 中移除该键并返回对应的值。
-    //      如果键不存在，remove 返回 None。
-    //      *key 是解引用（deref），将 &str（引用）转换为 str，因为 remove 接受 &str 类型。
-    //    - find_map: 遍历所有候选键，返回第一个 Some(值) 的结果。
-    //      即找到第一个存在的键就停止。
-    //
-    // 4. .and_then(|v| v.as_str().map(String::from)) — 如果找到了时间戳值：
-    //    - v.as_str(): 尝试将 serde_json::Value 转换为 &str（字符串切片）。
-    //      如果 Value 不是字符串类型（比如是数字），返回 None。
-    //    - .map(String::from): 如果是字符串，将其转换为 String（拥有所有权的字符串）。
-    //    - and_then: 类似于 map，但用于"扁平化"嵌套的 Option。
-    //      如果 as_str() 返回 None，整个链返回 None。
-    let timestamp = ["timestamp", "time", "ts", "@timestamp"]
-        .iter()
-        .find_map(|key| fields.remove(*key))
-        .and_then(|v| v.as_str().map(String::from));
-
-    // ─── 提取日志级别字段 ──────────────────────────────────────────────
-    // 与时间戳提取类似，但多了一步：将字符串解析为 LogLevel 枚举。
-    let level = ["level", "lvl", "severity"]
-        .iter()
-        .find_map(|key| fields.remove(*key))
-        .and_then(|v| v.as_str().map(String::from))
-        // .map(|s| s.parse::<LogLevel>(...)) — 尝试将字符串解析为 LogLevel 枚举。
-        // parse::<LogLevel> 中的 ::<LogLevel> 是泛型参数（turbofish 语法），
-        // 指定我们要将字符串解析为 LogLevel 类型。
-        //
-        // .unwrap_or_else(|e| match e {}) — 错误处理：
-        //   - 如果解析成功，直接返回 LogLevel 值。
-        //   - 如果解析失败（字符串不匹配任何已知的日志级别），执行闭包。
-        //   - |e| match e {}: 这个闭包接收解析错误 e，用 match e {} 进行"穷尽匹配"。
-        //     由于 LogLevel 的 parse 错误类型是一个空枚举（没有任何变体），
-        //     match e {} 意味着"这个分支永远不会执行"（unreachable）。
-        //     但实际上，如果 parse 失败，unwrap_or_else 不会执行这个闭包——
-        //     等等，这里有个细微之处：
-        //     unwrap_or_else 只在 Err 时执行闭包，但 match e {} 对空枚举是合法的
-        //     （因为空枚举没有任何可能的值，所以 match 是穷尽的）。
-        //     不过这里的实际效果是：如果 parse 失败，整个 .map() 返回 None
-        //     （因为 unwrap_or_else 返回的类型是 LogLevel，而空 match 不会有返回值）。
-        //
-        //     实际上更准确的解释：parse() 的错误类型是 Infallible（不可失败的），
-        //     即解析总是成功。所以 unwrap_or_else 永远不会被执行。
-        //     但即使如此，unwrap_or_else 的闭包也需要类型正确，match e {} 满足这一点。
         .map(|s| s.parse::<LogLevel>().unwrap_or_else(|e| match e {}));
 
-    // ─── 构建 LogEntry 并返回 ──────────────────────────────────────────
+    // serde_json::Map → HashMap：剩余字段转为 HashMap 存入 fields
-    // 此时 fields HashMap 中还剩下未被提取的字段（如 message、自定义字段等）。
+    let fields: HashMap<String, Value> = obj.into_iter().collect();
-    // timestamp 和 level 已经从 fields 中移除了（通过 remove）。
+
-    //
-    // Some(LogEntry { ... }) — 使用结构体字面量创建 LogEntry 实例，
-    // 并用 Some() 包裹表示"有值"。
     Some(LogEntry {
-        // line_number 设为 0，由调用者（如 parse_line_with_number）设置正确的值。
         line_number: 0,
-        // 原始行内容。
         raw_line,
-        // 时间戳（可能为 None，如果 JSON 中没有时间戳字段）。
         timestamp,
-        // 日志级别（可能为 None，如果 JSON 中没有级别字段）。
         level,
-        // 剩余的 JSON 字段（已移除 timestamp 和 level）。
         fields,
+        duplicate_keys,
     })
 }
 
@@ -317,6 +350,13 @@ mod tests {
         assert_eq!(parse_line(warn_line).unwrap().level, Some(LogLevel::Warn));
     }
 
+    #[test]
+    // Regression: level field with surrounding whitespace should still be recognized.
+    fn test_level_whitespace_in_json() {
+        let line = r#"{"level":" WARN ","message":"test"}"#;
+        assert_eq!(parse_line(line).unwrap().level, Some(LogLevel::Warn));
+    }
+
     #[test]
     // 测试：所有候选时间戳键名（timestamp, time, ts, @timestamp）都能被识别。
     fn test_timestamp_key_names() {
@@ -357,4 +397,180 @@ mod tests {
             assert_eq!(entry.level, Some(LogLevel::Info), "failed for key: {key}");
         }
     }
+
+    #[test]
+    // 测试：数字类型的 level 值不应被提取，应保留在 fields 中。
+    fn test_numeric_level_preserved_in_fields() {
+        let line = r#"{"level":30,"message":"hello"}"#;
+        let entry = parse_line(line).unwrap();
+        // level 不是字符串，应返回 None。
+        assert!(entry.level.is_none());
+        // 数字 level 应保留在 fields 中，不被静默丢弃。
+        assert_eq!(entry.fields.get("level"), Some(&Value::Number(30.into())));
+        // message 仍正常存在。
+        assert_eq!(
+            entry.fields.get("message"),
+            Some(&Value::String("hello".to_string()))
+        );
+    }
+
+    #[test]
+    // 测试：数字类型的 timestamp 值不应被提取，应保留在 fields 中。
+    fn test_numeric_timestamp_preserved_in_fields() {
+        let line = r#"{"timestamp":1718000000,"message":"hello"}"#;
+        let entry = parse_line(line).unwrap();
+        // timestamp 不是字符串，应返回 None。
+        assert!(entry.timestamp.is_none());
+        // 数字 timestamp 应保留在 fields 中。
+        assert_eq!(
+            entry.fields.get("timestamp"),
+            Some(&Value::Number(1718000000.into()))
+        );
+    }
+
+    #[test]
+    // 测试：当第一个候选键是数字时，应回退到下一个字符串类型的候选键。
+    fn test_fallback_to_string_key() {
+        let line = r#"{"level":30,"lvl":"INFO","message":"hello"}"#;
+        let entry = parse_line(line).unwrap();
+        // "level" 是数字，应跳过；"lvl" 是字符串，应成功提取。
+        assert_eq!(entry.level, Some(LogLevel::Info));
+        // 数字 "level" 保留在 fields 中。
+        assert_eq!(entry.fields.get("level"), Some(&Value::Number(30.into())));
+        // "lvl" 已被成功提取并从 fields 中移除。
+        assert!(entry.fields.get("lvl").is_none());
+    }
+
+    #[test]
+    // 测试：timestamp 的 fallback 行为 — 数字 timestamp 被保留，字符串 time 被提取。
+    fn test_timestamp_fallback_preserves_numeric() {
+        let line = r#"{"timestamp":1718000000,"time":"2024-01-01T00:00:00Z"}"#;
+        let entry = parse_line(line).unwrap();
+        // "timestamp" 是数字，跳过；"time" 是字符串，成功提取。
+        assert_eq!(entry.timestamp, Some("2024-01-01T00:00:00Z".to_string()));
+        // 数字 "timestamp" 保留在 fields 中。
+        assert_eq!(
+            entry.fields.get("timestamp"),
+            Some(&Value::Number(1718000000.into()))
+        );
+        // "time" 已被提取并从 fields 中移除。
+        assert!(entry.fields.get("time").is_none());
+    }
+
+    #[test]
+    fn test_bom_prefixed_json() {
+        let line = "\u{FEFF}{\"level\":\"INFO\",\"message\":\"hello\"}";
+        let entry = parse_line(line).unwrap();
+        assert_eq!(entry.level, Some(LogLevel::Info));
+        assert_eq!(
+            entry.fields.get("message"),
+            Some(&Value::String("hello".into()))
+        );
+    }
+
+    #[test]
+    fn test_bom_prefixed_detect() {
+        assert!(detect_json_log("\u{FEFF}{\"level\":\"INFO\"}"));
+    }
+
+    #[test]
+    fn test_bom_only_whitespace() {
+        assert!(parse_line("\u{FEFF}   ").is_none());
+    }
+
+    #[test]
+    fn test_bom_stripped_from_raw_line() {
+        let line = "\u{FEFF}{\"level\":\"INFO\",\"message\":\"hello\"}";
+        let entry = parse_line(line).unwrap();
+        assert_eq!(entry.raw_line, "{\"level\":\"INFO\",\"message\":\"hello\"}");
+    }
+
+    #[test]
+    fn test_internal_bom_not_stripped() {
+        let line = "{\"message\":\"\u{FEFF}hello\"}";
+        let entry = parse_line(line).unwrap();
+        assert_eq!(
+            entry.fields.get("message"),
+            Some(&Value::String("\u{FEFF}hello".into()))
+        );
+    }
+
+    // ─── 重复 key 检测测试 ──────────────────────────────────────────────
+
+    #[test]
+    fn test_no_duplicate_keys_normal_json() {
+        let line = r#"{"level":"INFO","message":"hello"}"#;
+        let entry = parse_line(line).unwrap();
+        assert!(entry.duplicate_keys.is_empty());
+    }
+
+    #[test]
+    fn test_duplicate_message_key_detected() {
+        let line = r#"{"level":"INFO","message":"first","message":"second"}"#;
+        let entry = parse_line(line).unwrap();
+        // last-wins: fields 中保留第二个值
+        assert_eq!(
+            entry.fields.get("message"),
+            Some(&Value::String("second".into()))
+        );
+        // 重复 key 记录中包含所有值
+        assert_eq!(entry.duplicate_keys.len(), 1);
+        assert_eq!(entry.duplicate_keys[0].key, "message");
+        assert_eq!(entry.duplicate_keys[0].values.len(), 2);
+        assert_eq!(
+            entry.duplicate_keys[0].values[0],
+            Value::String("first".into())
+        );
+        assert_eq!(
+            entry.duplicate_keys[0].values[1],
+            Value::String("second".into())
+        );
+    }
+
+    #[test]
+    fn test_duplicate_key_last_wins() {
+        let line = r#"{"msg":"a","msg":"b","msg":"c"}"#;
+        let entry = parse_line(line).unwrap();
+        // last-wins: 最终值是 "c"
+        assert_eq!(entry.fields.get("msg"), Some(&Value::String("c".into())));
+        // 三个值都被记录
+        assert_eq!(entry.duplicate_keys.len(), 1);
+        assert_eq!(entry.duplicate_keys[0].values.len(), 3);
+        assert_eq!(entry.duplicate_keys[0].values[0], Value::String("a".into()));
+        assert_eq!(entry.duplicate_keys[0].values[1], Value::String("b".into()));
+        assert_eq!(entry.duplicate_keys[0].values[2], Value::String("c".into()));
+    }
+
+    #[test]
+    fn test_multiple_different_duplicate_keys() {
+        let line = r#"{"a":"1","b":"2","a":"3","b":"4"}"#;
+        let entry = parse_line(line).unwrap();
+        assert_eq!(entry.duplicate_keys.len(), 2);
+        let dup_a = entry.duplicate_keys.iter().find(|d| d.key == "a").unwrap();
+        let dup_b = entry.duplicate_keys.iter().find(|d| d.key == "b").unwrap();
+        assert_eq!(dup_a.values.len(), 2);
+        assert_eq!(dup_b.values.len(), 2);
+    }
+
+    #[test]
+    fn test_duplicate_level_key_detected() {
+        let line = r#"{"level":"INFO","level":"ERROR","message":"hello"}"#;
+        let entry = parse_line(line).unwrap();
+        // last-wins: level 被提取为 ERROR
+        assert_eq!(entry.level, Some(LogLevel::Error));
+        // 重复 key 被记录
+        assert_eq!(entry.duplicate_keys.len(), 1);
+        assert_eq!(entry.duplicate_keys[0].key, "level");
+        assert_eq!(entry.duplicate_keys[0].values.len(), 2);
+    }
+
+    #[test]
+    fn test_duplicate_timestamp_key_detected() {
+        let line = r#"{"timestamp":"2024-01-01","timestamp":"2024-06-01"}"#;
+        let entry = parse_line(line).unwrap();
+        // last-wins: timestamp 提取为后者
+        assert_eq!(entry.timestamp, Some("2024-06-01".to_string()));
+        assert_eq!(entry.duplicate_keys.len(), 1);
+        assert_eq!(entry.duplicate_keys[0].key, "timestamp");
+    }
 }

crates/core/src/parser/level.rs

@@ -100,13 +100,21 @@ fn detect_level_from_text(line: &str) -> Option<LogLevel> {
     )
 }
 
+// ─── is_ident_char ─────────────────────────────────────────────────────────
+/// Whether a byte looks like an ASCII identifier continuation character
+/// (letter / digit / underscore). Log-level keywords must NOT be adjacent to
+/// such characters to count as a valid word boundary.
+fn is_ident_char(b: u8) -> bool {
+    b.is_ascii_alphanumeric() || b == b'_'
+}
+
 // ─── is_word_boundary ───────────────────────────────────────────────────────
-/// Check that the match at `start..start+len` is surrounded by non-alphabetic
+/// Check that the match at `start..start+len` is surrounded by non-identifier
 /// characters (or the string edge).
 fn is_word_boundary(text: &str, start: usize, len: usize) -> bool {
-    let before_ok = start == 0 || !text.as_bytes()[start - 1].is_ascii_alphabetic();
+    let before_ok = start == 0 || !is_ident_char(text.as_bytes()[start - 1]);
     let after_idx = start + len;
-    let after_ok = after_idx >= text.len() || !text.as_bytes()[after_idx].is_ascii_alphabetic();
+    let after_ok = after_idx >= text.len() || !is_ident_char(text.as_bytes()[after_idx]);
     before_ok && after_ok
 }
 
@@ -209,4 +217,36 @@ mod tests {
         let line = format!("{prefix} ERROR something");
         assert_eq!(detect_level(&line), Some(LogLevel::Error));
     }
+
+    #[test]
+    fn test_boundary_rejects_trailing_digits() {
+        assert_eq!(detect_level("ERROR123"), None);
+        assert_eq!(detect_level("WARN2: bad"), None);
+        assert_eq!(detect_level("ERR2"), None);
+    }
+
+    #[test]
+    fn test_boundary_rejects_underscore() {
+        assert_eq!(detect_level("INFO_foo"), None);
+        assert_eq!(detect_level("DBG_value=5"), None);
+    }
+
+    #[test]
+    fn test_boundary_rejects_leading_digits_and_underscore() {
+        assert_eq!(detect_level("123ERROR: fail"), None);
+        assert_eq!(detect_level("foo_ERROR: fail"), None);
+        assert_eq!(detect_level("1WRN"), None);
+    }
+
+    #[test]
+    fn test_boundary_accepts_valid_suffixes() {
+        assert_eq!(detect_level("ERROR: fail"), Some(LogLevel::Error));
+        assert_eq!(detect_level("[ERROR] fail"), Some(LogLevel::Error));
+        assert_eq!(detect_level("ERROR fail"), Some(LogLevel::Error));
+    }
+
+    #[test]
+    fn test_boundary_camel_case_regression() {
+        assert_eq!(detect_level("errorLevel"), None);
+    }
 }

crates/core/src/types.rs

@@ -71,13 +71,17 @@ impl FromStr for LogLevel {
     // 接收一个字符串切片 &str，返回 Result<LogLevel, Infallible>。
     // 由于 Err 类型是 Infallible，实际上返回值总是 Ok(LogLevel)。
     fn from_str(s: &str) -> Result<Self, Self::Err> {
-        // `s.to_uppercase()` — 将字符串转换为大写，实现不区分大小写的匹配。
+        // `s.trim()` — 去除字符串前后的 Unicode 空白字符。
+        // 例如 " WARN " → "WARN"，"\tINFO\n" → "INFO"。
+        let trimmed = s.trim();
+
+        // `trimmed.to_uppercase()` — 将 trimmed 后的字符串转换为大写，实现不区分大小写的匹配。
         // 例如 "info"、"Info"、"INFO" 都会被转换为 "INFO"。
         // 返回一个新的 String（堆分配）。
         //
         // `.as_str()` — 将 String 转换回 &str（字符串切片引用）。
         // 因为 match 需要匹配 &str 而不是 String。
-        match s.to_uppercase().as_str() {
+        match trimmed.to_uppercase().as_str() {
             // `|` 在 match 分支中表示"或"（multiple patterns）。
             // "ERROR" | "ERR" | "SEVERE" | "FATAL" 都匹配到 LogLevel::Error。
             "ERROR" | "ERR" | "SEVERE" | "FATAL" => Ok(LogLevel::Error),
@@ -86,9 +90,9 @@ impl FromStr for LogLevel {
             "DEBUG" | "DBG" => Ok(LogLevel::Debug),
             "TRACE" | "TRC" => Ok(LogLevel::Trace),
             // `_` 是通配符，匹配所有未被上面分支捕获的值。
-            // 对于未知级别，包装为 Unknown 并保存原始字符串。
+            // 对于未知级别，包装为 Unknown 并保存 trimmed 后的字符串。
-            // s.to_string() 将 &str 转换为 String（注意这里用原始的 s，不是大写后的）。
+            // s.to_string() 将 &str 转换为 String（注意这里用 trimmed，不是原始 s）。
-            _ => Ok(LogLevel::Unknown(s.to_string())),
+            _ => Ok(LogLevel::Unknown(trimmed.to_string())),
         }
     }
 }
@@ -119,6 +123,18 @@ impl fmt::Display for LogLevel {
 }
 
 // ─── LogEntry 结构体 ────────────────────────────────────────────────────────
+/// 记录 JSON 日志中出现的重复 key 信息
+///
+/// 当 JSON 对象中同一个 key 出现多次时，serde_json 默认 last-wins（后值覆盖前值），
+/// 前面的值会静默丢失。此结构记录所有重复出现的 key 及其全部值，供 UI 展示警告。
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
+pub struct DuplicateKey {
+    /// 重复的 key 名称
+    pub key: String,
+    /// 该 key 出现的所有值（按出现顺序排列，最后一个值是 fields 中的最终值）
+    pub values: Vec<Value>,
+}
+
 /// 一行解析后的日志
 ///
 /// 表示日志文件中经过解析器处理后的一行内容。
@@ -144,6 +160,9 @@ pub struct LogEntry {
     /// HashMap<String, Value> 是一个字典，键是字段名，值是 JSON 值。
     /// 例如 {"message": "hello", "request_id": "abc123"}。
     pub fields: HashMap<String, Value>,
+
+    /// JSON 中重复出现的 key 记录（正常日志为空 Vec）
+    pub duplicate_keys: Vec<DuplicateKey>,
 }
 
 // ─── SearchResult 结构体 ────────────────────────────────────────────────────
@@ -288,6 +307,31 @@ mod tests {
         );
     }
 
+    #[test]
+    fn test_from_str_whitespace_trimmed() {
+        assert_eq!(" WARN ".parse::<LogLevel>(), Ok(LogLevel::Warn));
+        assert_eq!("\tINFO".parse::<LogLevel>(), Ok(LogLevel::Info));
+        assert_eq!("ERROR\n".parse::<LogLevel>(), Ok(LogLevel::Error));
+        assert_eq!("  debug  ".parse::<LogLevel>(), Ok(LogLevel::Debug));
+        assert_eq!("\tTRACE\t".parse::<LogLevel>(), Ok(LogLevel::Trace));
+    }
+
+    #[test]
+    fn test_from_str_whitespace_unknown_trimmed() {
+        // Unknown stores the trimmed value, not the original.
+        assert_eq!(
+            " CUSTOM ".parse::<LogLevel>(),
+            Ok(LogLevel::Unknown("CUSTOM".into()))
+        );
+        // Pure whitespace becomes Unknown("").
+        assert_eq!("   ".parse::<LogLevel>(), Ok(LogLevel::Unknown("".into())));
+        // Internal whitespace is NOT collapsed.
+        assert_eq!(
+            "W ARN".parse::<LogLevel>(),
+            Ok(LogLevel::Unknown("W ARN".into()))
+        );
+    }
+
     #[test]
     // 测试：LogLevel 的 Display 输出格式是否正确。
     fn test_display_output() {
@@ -319,6 +363,7 @@ mod tests {
             timestamp: Some("2024-01-01T00:00:00".to_string()),
             level: Some(LogLevel::Info),
             fields,
+            duplicate_keys: vec![],
         };
 
         // 逐字段验证。

crates/core/src/watcher/file_watcher.rs

@@ -1,7 +1,7 @@
 use std::path::{Path, PathBuf};
 use std::sync::{Arc, Mutex};
 
-use crossbeam_channel::{bounded, Receiver, Sender};
+use crossbeam_channel::{Receiver, Sender, bounded};
 use notify::{Event, EventKind, RecommendedWatcher, RecursiveMode, Watcher};
 
 use crate::error::Result;
@@ -13,6 +13,7 @@ pub enum FileEvent {
     Truncated { new_size: u64 },
     Rotated { new_inode: u64 },
     Removed,
+    WatcherError { message: String },
 }
 
 // ─── get_inode ──────────────────────────────────────────────────────────────
@@ -34,6 +35,43 @@ struct WatchState {
     last_inode: u64,
 }
 
+fn process_event(event: Event, watch_path: &Path, state: &mut WatchState) -> Option<FileEvent> {
+    if !event.paths.iter().any(|p| p == watch_path) {
+        return None;
+    }
+
+    match event.kind {
+        EventKind::Modify(_) | EventKind::Create(_) | EventKind::Any => {}
+        EventKind::Remove(_) => return Some(FileEvent::Removed),
+        _ => return None,
+    }
+
+    let current_inode = get_inode(watch_path).unwrap_or(0);
+    let current_size = std::fs::metadata(watch_path).map(|m| m.len()).unwrap_or(0);
+
+    if current_inode != 0 && state.last_inode != 0 && current_inode != state.last_inode {
+        state.last_inode = current_inode;
+        state.last_size = current_size;
+        Some(FileEvent::Rotated {
+            new_inode: current_inode,
+        })
+    } else if current_size > state.last_size {
+        state.last_size = current_size;
+        state.last_inode = current_inode;
+        Some(FileEvent::Appended {
+            new_size: current_size,
+        })
+    } else if current_size < state.last_size {
+        state.last_size = current_size;
+        state.last_inode = current_inode;
+        Some(FileEvent::Truncated {
+            new_size: current_size,
+        })
+    } else {
+        None
+    }
+}
+
 // ─── FileWatcher ────────────────────────────────────────────────────────────
 pub struct FileWatcher {
     rx: Receiver<FileEvent>,
@@ -56,50 +94,22 @@ impl FileWatcher {
             notify::recommended_watcher(move |res: std::result::Result<Event, notify::Error>| {
                 let event = match res {
                     Ok(e) => e,
-                    Err(_) => return,
+                    Err(error) => {
+                        let _ = tx.try_send(FileEvent::WatcherError {
+                            message: error.to_string(),
+                        });
+                        return;
+                    }
                 };
 
-                match event.kind {
-                    EventKind::Modify(_) | EventKind::Create(_) | EventKind::Any => {}
-                    EventKind::Remove(_) => {
-                        let _ = tx.send(FileEvent::Removed);
-                        return;
-                    }
-                    _ => return,
-                }
-
-                if !event.paths.iter().any(|p| p == &watch_path) {
-                    return;
-                }
-
-                let current_inode = get_inode(&watch_path).unwrap_or(0);
-                let current_size = std::fs::metadata(&watch_path).map(|m| m.len()).unwrap_or(0);
-
                 let mut st = state.lock().unwrap_or_else(|poison| {
                     // Recover from poisoned mutex — state only tracks last_size
                     // and last_inode for event dedup. Stale values at worst
                     // cause a duplicate event, which is harmless.
                     poison.into_inner()
                 });
-
+                if let Some(fe) = process_event(event, &watch_path, &mut st) {
-                if current_inode != 0 && st.last_inode != 0 && current_inode != st.last_inode {
+                    let _ = tx.try_send(fe);
-                    let _ = tx.send(FileEvent::Rotated {
-                        new_inode: current_inode,
-                    });
-                    st.last_inode = current_inode;
-                    st.last_size = current_size;
-                } else if current_size > st.last_size {
-                    let _ = tx.send(FileEvent::Appended {
-                        new_size: current_size,
-                    });
-                    st.last_size = current_size;
-                    st.last_inode = current_inode;
-                } else if current_size < st.last_size {
-                    let _ = tx.send(FileEvent::Truncated {
-                        new_size: current_size,
-                    });
-                    st.last_size = current_size;
-                    st.last_inode = current_inode;
                 }
             })?;
 
@@ -144,6 +154,56 @@ mod tests {
         events
     }
 
+    #[test]
+    fn test_remove_wrong_path_ignored() {
+        let dir = tempfile::tempdir().expect("create temp dir");
+        let watched = dir.path().join("watched.log");
+        let other = dir.path().join("other.log");
+        std::fs::write(&watched, b"hello\n").expect("write watched");
+        std::fs::write(&other, b"other\n").expect("write other");
+
+        let mut state = WatchState {
+            last_size: 6,
+            last_inode: get_inode(&watched).unwrap_or(0),
+        };
+
+        let event = Event {
+            kind: EventKind::Remove(notify::event::RemoveKind::File),
+            paths: vec![other.clone()],
+            attrs: Default::default(),
+        };
+
+        let result = process_event(event, &watched, &mut state);
+        assert_eq!(
+            result, None,
+            "Remove for non-watched path should be ignored"
+        );
+    }
+
+    #[test]
+    fn test_remove_correct_path_emits_removed() {
+        let dir = tempfile::tempdir().expect("create temp dir");
+        let watched = dir.path().join("watched.log");
+        std::fs::write(&watched, b"hello\n").expect("write watched");
+
+        let mut state = WatchState {
+            last_size: 6,
+            last_inode: get_inode(&watched).unwrap_or(0),
+        };
+
+        let event = Event {
+            kind: EventKind::Remove(notify::event::RemoveKind::File),
+            paths: vec![watched.clone()],
+            attrs: Default::default(),
+        };
+
+        let result = process_event(event, &watched, &mut state);
+        assert!(
+            matches!(result, Some(FileEvent::Removed)),
+            "Remove for watched path should emit Removed"
+        );
+    }
+
     #[test]
     fn test_watcher_append() {
         let dir = tempfile::tempdir().expect("create temp dir");
@@ -232,5 +292,19 @@ mod tests {
 
         let d = FileEvent::Rotated { new_inode: 42 };
         assert_ne!(a, d);
+
+        let e1 = FileEvent::WatcherError {
+            message: "io error".into(),
+        };
+        let e2 = FileEvent::WatcherError {
+            message: "io error".into(),
+        };
+        assert_eq!(e1, e2);
+        assert_ne!(
+            e1,
+            FileEvent::WatcherError {
+                message: "other".into()
+            }
+        );
     }
 }

crates/tui/Cargo.toml

@@ -11,6 +11,7 @@ anyhow.workspace = true
 log-viewer-core.workspace = true
 serde_json.workspace = true
 crossbeam-channel.workspace = true
+unicode-width.workspace = true
 
 [dev-dependencies]
 tempfile = { workspace = true }

crates/tui/src/app/input.rs

@@ -0,0 +1,138 @@
+use std::time::Instant;
+
+use crossterm::event::{KeyCode, KeyEvent, KeyEventKind, KeyModifiers};
+
+use super::{App, AppMode};
+
+impl App {
+    pub fn handle_key(&mut self, key: KeyEvent) {
+        let should_handle = match key.kind {
+            KeyEventKind::Press => true,
+            KeyEventKind::Repeat => self.is_repeatable_key(&key),
+            KeyEventKind::Release => false,
+        };
+
+        if !should_handle {
+            return;
+        }
+
+        match self.mode {
+            AppMode::Normal => self.handle_normal_key(key),
+            AppMode::Settings => self.handle_settings_key(key),
+        }
+    }
+
+    /// Keys that should auto-repeat when held (scroll/navigation only).
+    fn is_repeatable_key(&self, key: &KeyEvent) -> bool {
+        let plain = key.modifiers.is_empty();
+        let ctrl = key.modifiers == KeyModifiers::CONTROL;
+
+        match self.mode {
+            AppMode::Normal => {
+                (plain
+                    && matches!(
+                        key.code,
+                        KeyCode::Char('j')
+                            | KeyCode::Down
+                            | KeyCode::Char('k')
+                            | KeyCode::Up
+                            | KeyCode::PageDown
+                            | KeyCode::PageUp
+                    ))
+                    || (ctrl
+                        && matches!(
+                            key.code,
+                            KeyCode::Char('d')
+                                | KeyCode::Char('u')
+                                | KeyCode::Char('f')
+                                | KeyCode::Char('b')
+                        ))
+            }
+            AppMode::Settings => {
+                plain
+                    && matches!(
+                        key.code,
+                        KeyCode::Char('j')
+                            | KeyCode::Down
+                            | KeyCode::Char('k')
+                            | KeyCode::Up
+                            | KeyCode::Left
+                            | KeyCode::Right
+                    )
+            }
+        }
+    }
+
+    fn handle_normal_key(&mut self, key: KeyEvent) {
+        match key.code {
+            KeyCode::Char('q') | KeyCode::Esc => {
+                self.should_quit = true;
+                self.last_g_press = None;
+            }
+            KeyCode::Char('j') | KeyCode::Down => {
+                self.scroll_down_line();
+                self.last_g_press = None;
+            }
+            KeyCode::Char('k') | KeyCode::Up => {
+                self.scroll_up_line();
+                self.last_g_press = None;
+            }
+            KeyCode::Char('d') if key.modifiers.contains(KeyModifiers::CONTROL) => {
+                self.scroll_down_half_page();
+                self.last_g_press = None;
+            }
+            KeyCode::Char('u') if key.modifiers.contains(KeyModifiers::CONTROL) => {
+                self.scroll_up_half_page();
+                self.last_g_press = None;
+            }
+            KeyCode::Char('f') if key.modifiers.contains(KeyModifiers::CONTROL) => {
+                self.scroll_down_page();
+                self.last_g_press = None;
+            }
+            KeyCode::PageDown => {
+                self.scroll_down_page();
+                self.last_g_press = None;
+            }
+            KeyCode::Char('b') if key.modifiers.contains(KeyModifiers::CONTROL) => {
+                self.scroll_up_page();
+                self.last_g_press = None;
+            }
+            KeyCode::PageUp => {
+                self.scroll_up_page();
+                self.last_g_press = None;
+            }
+            KeyCode::Char('G') | KeyCode::End => {
+                self.scroll_to_bottom();
+                self.last_g_press = None;
+            }
+            KeyCode::Char('g') => {
+                if let Some(instant) = self.last_g_press
+                    && instant.elapsed().as_millis() < 500
+                {
+                    self.scroll_to_top();
+                    self.last_g_press = None;
+                    return;
+                }
+                self.last_g_press = Some(Instant::now());
+            }
+            KeyCode::Home => {
+                self.scroll_to_top();
+                self.last_g_press = None;
+            }
+            KeyCode::Tab => {
+                self.toggle_json_format();
+                self.last_g_press = None;
+            }
+            KeyCode::Char('s') | KeyCode::Char('S')
+                if !key.modifiers.contains(KeyModifiers::CONTROL) =>
+            {
+                self.settings_draft = self.color_config.clone();
+                self.settings_error = None;
+                self.mode = AppMode::Settings;
+            }
+            _ => {
+                self.last_g_press = None;
+            }
+        }
+    }
+}

crates/tui/src/app/loading.rs

@@ -0,0 +1,169 @@
+use std::path::Path;
+
+use log_viewer_core::io::progressive_reader::{
+    IndexerMessage, ProgressiveFileReader, spawn_indexer,
+};
+use log_viewer_core::watcher::file_watcher::FileWatcher;
+
+use super::{App, AppLoadingState};
+
+impl App {
+    pub fn load_file(&mut self, path: &str) -> anyhow::Result<()> {
+        // ── Phase 1: Pure computation, no mutation of self ──────────
+        // If any step fails and returns ?, self remains completely untouched,
+        // preserving the old file's watcher, loading_state, and file_path.
+        let mut pfr =
+            ProgressiveFileReader::open(Path::new(path)).map_err(|e| anyhow::anyhow!("{e}"))?;
+
+        let new_loading_state = if pfr.is_sampling() {
+            // Cache miss: spawn background indexer
+            let (cancel_tx, cancel_rx) = crossbeam_channel::bounded(1);
+            let generation = pfr.generation();
+            let indexer_rx =
+                spawn_indexer(pfr.path().to_path_buf(), generation, 80, false, cancel_rx);
+            pfr = ProgressiveFileReader::with_channels(
+                Path::new(path),
+                cancel_tx,
+                indexer_rx,
+                generation,
+            )
+            .map_err(|e| anyhow::anyhow!("{e}"))?;
+
+            let estimated = pfr.line_count() as u64;
+            AppLoadingState::Loading {
+                reader: pfr,
+                estimated_lines: estimated,
+                progress_percent: 0.0,
+            }
+        } else {
+            // Cache hit: Ready state
+            AppLoadingState::Ready { reader: pfr }
+        };
+
+        let new_watcher = FileWatcher::watch(Path::new(path)).ok();
+
+        // ── Phase 2: Commit — swap self to new state ───────────────
+        // SAFETY: Do NOT add any fallible operations (with ?) below this point.
+        // The old file_watcher and loading_state are dropped here, cancelling
+        // any background indexer for the previous file.
+        self.file_watcher = new_watcher;
+        self.loading_state = new_loading_state;
+        self.file_path = Some(path.to_string());
+
+        // Reset UI state for the new file
+        self.cursor_line = 0;
+        self.v_offset = 0;
+        self.v_sub_offset = 0;
+        self.cursor_sub_offset = 0;
+        self.viewport_cache.invalidate();
+        self.last_g_press = None;
+        self.json_format = false;
+        self.mode = super::AppMode::Normal;
+        self.reload_after_loading = false;
+        Ok(())
+    }
+
+    /// Poll the background indexer for progress/completion.
+    /// Transitions Loading → Ready when indexing completes.
+    /// Must be called every frame in the event loop.
+    pub fn poll_background_indexer(&mut self) {
+        // Poll visual height rebuild (Ready state only).
+        // Recalibrate v_offset only when VHI actually changed.
+        if let AppLoadingState::Ready { reader } = &mut self.loading_state
+            && let Some(index) = reader.poll_visual_height_rebuild()
+        {
+            if let log_viewer_core::io::progressive_reader::ReaderState::Ready {
+                visual_height_index,
+                ..
+            } = &mut reader.state
+            {
+                *visual_height_index = Some(index);
+            }
+
+            let logical_top = self.v_offset.min(self.total_lines().saturating_sub(1));
+            let sub = self.v_sub_offset;
+            let first_visual = self.cursor_to_first_visual_row(logical_top);
+            let line_height = self
+                .get_visual_height_index()
+                .map_or(1, |idx| idx.visual_height_of_line(logical_top));
+            self.v_offset = first_visual.saturating_add(sub.min(line_height.saturating_sub(1)));
+            self.v_sub_offset = 0;
+            self.clamp_cursor_sub_offset_no_vhi();
+            self.clamp_v_offset();
+            self.viewport_cache.invalidate();
+        }
+
+        // Poll main indexer (Loading state)
+        let old_state = std::mem::replace(&mut self.loading_state, AppLoadingState::Empty);
+
+        if let AppLoadingState::Loading {
+            mut reader,
+            estimated_lines,
+            mut progress_percent,
+        } = old_state
+        {
+            if let Some(msg) = reader.poll_indexer() {
+                match msg {
+                    IndexerMessage::Progress { percent, .. } => {
+                        progress_percent = percent;
+                        self.loading_state = AppLoadingState::Loading {
+                            reader,
+                            estimated_lines,
+                            progress_percent,
+                        };
+                    }
+                    IndexerMessage::Complete {
+                        reader: fr,
+                        visual_height_index,
+                        ..
+                    } => {
+                        let saved_cursor = self.cursor_line;
+
+                        reader.set_ready(fr, visual_height_index);
+                        self.loading_state = AppLoadingState::Ready { reader };
+                        self.viewport_cache.invalidate();
+
+                        // Clamp cursor if exact count < estimated
+                        self.cursor_line = saved_cursor.min(self.total_lines().saturating_sub(1));
+
+                        // Loading uses 1:1 logical-line offsets; Ready uses visual-row
+                        // offsets derived from the prefix-sum index.  Recompute v_offset
+                        // so the same logical line stays visible (falls back to 1:1 when
+                        // the index is absent, which is the case right after invalidate).
+                        self.v_offset = self.cursor_to_first_visual_row(self.cursor_line);
+                        self.clamp_v_offset();
+                        self.v_sub_offset = 0;
+                        self.cursor_sub_offset = 0;
+
+                        // Gutter width changes (~N → N) shift content_width, so any
+                        // VisualHeightIndex built with the old width is stale.
+                        let (new_offset, new_sub) = self.rebase_offset_for_invalidate();
+                        if let AppLoadingState::Ready { reader } = &mut self.loading_state {
+                            reader.invalidate_visual_height_index();
+                        }
+                        self.v_offset = new_offset;
+                        self.v_sub_offset = new_sub;
+                        self.cursor_sub_offset = 0;
+
+                        if self.reload_after_loading {
+                            self.reload_after_loading = false;
+                            self.reload_ready_reader();
+                        }
+                    }
+                    IndexerMessage::Error { message, .. } => {
+                        self.loading_state = AppLoadingState::Error(message);
+                        self.reload_after_loading = false;
+                    }
+                }
+            } else {
+                self.loading_state = AppLoadingState::Loading {
+                    reader,
+                    estimated_lines,
+                    progress_percent,
+                };
+            }
+        } else {
+            self.loading_state = old_state;
+        }
+    }
+}

crates/tui/src/app/query.rs

@@ -0,0 +1,209 @@
+use log_viewer_core::config::ColorConfig;
+use log_viewer_core::io::progressive_reader::VisualHeightIndex;
+
+use super::viewport_cache::gutter_width_for;
+use super::{App, AppLoadingState, AppMode, ViewportRenderRow};
+
+impl App {
+    #[allow(dead_code)]
+    pub fn get_line(&self, idx: usize) -> Option<String> {
+        match &self.loading_state {
+            AppLoadingState::Ready { reader } => reader.get_line(idx),
+            AppLoadingState::Loading { reader, .. } => reader.get_line(idx),
+            _ => None,
+        }
+    }
+
+    #[allow(dead_code)]
+    pub fn file_name(&self) -> Option<&str> {
+        self.file_path
+            .as_ref()
+            .and_then(|p| std::path::Path::new(p).file_name().and_then(|n| n.to_str()))
+    }
+
+    pub fn total_lines(&self) -> usize {
+        match &self.loading_state {
+            AppLoadingState::Ready { reader } => reader.line_count(),
+            AppLoadingState::Loading {
+                reader,
+                estimated_lines,
+                ..
+            } => {
+                // Use estimated total lines (not sampled_line_count) so the user can
+                // scroll freely during indexing. get_line() incrementally scans
+                // forward on demand, so lines beyond the initial 64KB are still
+                // accessible. The .max() guards against under-estimates.
+                (*estimated_lines as usize).max(reader.sampled_line_count())
+            }
+            _ => 0,
+        }
+    }
+
+    pub fn is_loaded(&self) -> bool {
+        matches!(
+            self.loading_state,
+            AppLoadingState::Ready { .. } | AppLoadingState::Loading { .. }
+        )
+    }
+
+    pub fn is_loading(&self) -> bool {
+        matches!(self.loading_state, AppLoadingState::Loading { .. })
+    }
+
+    pub fn is_error(&self) -> bool {
+        matches!(self.loading_state, AppLoadingState::Error(_))
+    }
+
+    pub(super) fn get_visual_height_index(&self) -> Option<&VisualHeightIndex> {
+        match &self.loading_state {
+            AppLoadingState::Ready { reader } => match &reader.state {
+                log_viewer_core::io::progressive_reader::ReaderState::Ready {
+                    visual_height_index,
+                    ..
+                } => visual_height_index.as_ref(),
+                _ => None,
+            },
+            _ => None,
+        }
+    }
+
+    pub fn error_message(&self) -> Option<&str> {
+        match &self.loading_state {
+            AppLoadingState::Error(msg) => Some(msg),
+            _ => None,
+        }
+    }
+
+    pub fn loading_progress(&self) -> Option<f64> {
+        match &self.loading_state {
+            AppLoadingState::Loading {
+                progress_percent, ..
+            } => Some(*progress_percent),
+            _ => None,
+        }
+    }
+
+    pub fn estimated_lines(&self) -> Option<u64> {
+        match &self.loading_state {
+            AppLoadingState::Loading {
+                estimated_lines, ..
+            } => Some(*estimated_lines),
+            _ => None,
+        }
+    }
+
+    pub(crate) fn mode(&self) -> AppMode {
+        self.mode
+    }
+
+    pub(crate) fn settings_error(&self) -> Option<&str> {
+        self.settings_error.as_deref()
+    }
+
+    pub(crate) fn settings_draft(&self) -> &ColorConfig {
+        &self.settings_draft
+    }
+
+    pub(crate) fn settings_cursor(&self) -> usize {
+        self.settings_cursor
+    }
+
+    pub(crate) fn color_config(&self) -> &ColorConfig {
+        &self.color_config
+    }
+
+    pub(crate) fn cursor_line(&self) -> usize {
+        self.cursor_line
+    }
+
+    pub(crate) fn cursor_sub_offset(&self) -> usize {
+        self.cursor_sub_offset
+    }
+
+    #[allow(dead_code)]
+    pub(crate) fn content_width(&self) -> u16 {
+        self.content_width
+    }
+
+    #[allow(dead_code)]
+    pub(crate) fn content_height(&self) -> u16 {
+        self.content_height
+    }
+
+    pub(crate) fn set_content_area(&mut self, width: u16, height: u16) {
+        self.content_width = width;
+        self.content_height = height;
+    }
+
+    pub(crate) fn viewport_rows(
+        &self,
+        start_logical: usize,
+        offset_in_line: usize,
+        available_rows: usize,
+    ) -> Vec<ViewportRenderRow<'_>> {
+        let mut rows = Vec::new();
+
+        for (entry_idx, entry) in self.viewport_cache.entries.iter().enumerate() {
+            let logical_line = self.viewport_cache.logical_start + entry_idx;
+            let start_row = if logical_line == start_logical {
+                offset_in_line
+            } else {
+                0
+            };
+
+            for (visual_row, text) in entry.wrapped_rows.iter().enumerate().skip(start_row) {
+                if rows.len() >= available_rows {
+                    return rows;
+                }
+                rows.push(ViewportRenderRow {
+                    logical_line,
+                    visual_row,
+                    text,
+                    level: entry.level.as_ref(),
+                });
+            }
+        }
+
+        rows
+    }
+
+    #[cfg(test)]
+    pub(crate) fn enter_settings_mode_for_test(&mut self) {
+        self.mode = AppMode::Settings;
+        self.settings_draft = self.color_config.clone();
+    }
+
+    #[cfg(test)]
+    pub(crate) fn set_cursor_for_test(&mut self, line: usize, sub_offset: usize) {
+        self.cursor_line = line;
+        self.cursor_sub_offset = sub_offset;
+    }
+
+    #[cfg(test)]
+    pub fn set_error_state(&mut self, msg: impl Into<String>) {
+        self.loading_state = AppLoadingState::Error(msg.into());
+    }
+
+    #[allow(dead_code)]
+    pub fn file_size(&self) -> u64 {
+        match &self.loading_state {
+            AppLoadingState::Ready { reader } => reader.reader().map_or(0, |r| r.file_size()),
+            AppLoadingState::Loading { reader, .. } => reader.reader().map_or(0, |r| r.file_size()),
+            _ => 0,
+        }
+    }
+
+    /// MUST match the renderer's `gutter_width` formula in `ui.rs::render_content`.
+    pub(crate) fn gutter_width(&self) -> usize {
+        gutter_width_for(self.total_lines(), self.is_loading())
+    }
+
+    pub(super) fn get_content_width(&self) -> usize {
+        let total = if self.content_width > 0 {
+            self.content_width as usize
+        } else {
+            80
+        };
+        total.saturating_sub(gutter_width_for(self.total_lines(), self.is_loading()))
+    }
+}

crates/tui/src/app/scroll.rs

@@ -0,0 +1,359 @@
+use super::App;
+
+impl App {
+    fn line_visual_height(&self, line: usize, width: usize) -> usize {
+        if line >= self.total_lines() {
+            return 1;
+        }
+        self.compute_visual_height(line, width).max(1)
+    }
+
+    fn advance_visual_pos(
+        &self,
+        mut line: usize,
+        mut sub: usize,
+        mut n: usize,
+        width: usize,
+    ) -> (usize, usize) {
+        let last = self.total_lines().saturating_sub(1);
+        while n > 0 && line < last {
+            let h = self.line_visual_height(line, width);
+            let remaining_in_line = h.saturating_sub(sub);
+            if n < remaining_in_line {
+                sub += n;
+                return (line, sub);
+            }
+            n -= remaining_in_line;
+            line += 1;
+            sub = 0;
+        }
+        if line >= last {
+            let h = self.line_visual_height(last, width);
+            sub = (sub + n).min(h.saturating_sub(1));
+            line = last;
+        }
+        (line, sub)
+    }
+
+    fn retreat_visual_pos(
+        &self,
+        mut line: usize,
+        mut sub: usize,
+        mut n: usize,
+        width: usize,
+    ) -> (usize, usize) {
+        while n > 0 {
+            if sub >= n {
+                sub -= n;
+                return (line, sub);
+            }
+            n -= sub + 1;
+            if line == 0 {
+                return (0, 0);
+            }
+            line -= 1;
+            sub = self.line_visual_height(line, width).saturating_sub(1);
+        }
+        (line, sub)
+    }
+
+    pub(super) fn clamp_cursor_sub_offset_no_vhi(&mut self) {
+        if !self.is_loaded() || self.total_lines() == 0 {
+            return;
+        }
+        let width = self.get_content_width();
+        if width == 0 {
+            return;
+        }
+        let h = self.line_visual_height(self.cursor_line, width);
+        if self.cursor_sub_offset >= h {
+            self.cursor_sub_offset = h.saturating_sub(1);
+        }
+    }
+
+    pub(super) fn ensure_cursor_visible_no_vhi(&mut self) {
+        if !self.is_loaded() || self.total_lines() == 0 || self.content_height == 0 {
+            return;
+        }
+        let width = self.get_content_width();
+        if width == 0 {
+            return;
+        }
+        let content_h = self.content_height as usize;
+        let cur_line = self.cursor_line;
+        let cur_sub = self.cursor_sub_offset;
+        let v_top_line = self.v_offset;
+        let v_top_sub = self.v_sub_offset;
+
+        if cur_line < v_top_line || (cur_line == v_top_line && cur_sub < v_top_sub) {
+            self.v_offset = cur_line;
+            self.v_sub_offset = cur_sub;
+            return;
+        }
+
+        let mut line = v_top_line;
+        let mut sub = v_top_sub;
+        let mut walked = 0usize;
+        let total = self.total_lines();
+        while walked < content_h {
+            if line == cur_line && sub == cur_sub {
+                return;
+            }
+            let h = self.line_visual_height(line, width);
+            if sub + 1 < h {
+                sub += 1;
+            } else if line + 1 < total {
+                line += 1;
+                sub = 0;
+            } else {
+                return;
+            }
+            walked += 1;
+        }
+
+        let target_distance = content_h.saturating_sub(1);
+        let (new_line, new_sub) =
+            self.retreat_visual_pos(cur_line, cur_sub, target_distance, width);
+        self.v_offset = new_line;
+        self.v_sub_offset = new_sub;
+    }
+
+    pub fn scroll_down_line(&mut self) {
+        if !self.is_loaded() || self.total_lines() == 0 {
+            return;
+        }
+
+        if let Some(index) = self.get_visual_height_index() {
+            let last = self.total_lines().saturating_sub(1);
+            let current_height = index.visual_height_of_line(self.cursor_line);
+            if self.cursor_sub_offset + 1 < current_height {
+                self.cursor_sub_offset += 1;
+            } else if self.cursor_line < last {
+                self.cursor_line += 1;
+                self.cursor_sub_offset = 0;
+            }
+            self.ensure_cursor_visible();
+        } else {
+            let width = self.get_content_width();
+            if width > 0 && self.total_lines() > 0 {
+                let (new_line, new_sub) =
+                    self.advance_visual_pos(self.cursor_line, self.cursor_sub_offset, 1, width);
+                self.cursor_line = new_line;
+                self.cursor_sub_offset = new_sub;
+                self.ensure_cursor_visible_no_vhi();
+            }
+        }
+    }
+
+    pub fn scroll_up_line(&mut self) {
+        if !self.is_loaded() || self.total_lines() == 0 {
+            return;
+        }
+
+        if let Some(index) = self.get_visual_height_index() {
+            if self.cursor_sub_offset > 0 {
+                self.cursor_sub_offset -= 1;
+            } else if self.cursor_line > 0 {
+                let previous_line = self.cursor_line - 1;
+                let previous_line_last_sub =
+                    index.visual_height_of_line(previous_line).saturating_sub(1);
+                self.cursor_line = previous_line;
+                self.cursor_sub_offset = previous_line_last_sub;
+            }
+            self.ensure_cursor_visible();
+        } else {
+            let width = self.get_content_width();
+            if width > 0 {
+                let (new_line, new_sub) =
+                    self.retreat_visual_pos(self.cursor_line, self.cursor_sub_offset, 1, width);
+                self.cursor_line = new_line;
+                self.cursor_sub_offset = new_sub;
+                self.ensure_cursor_visible_no_vhi();
+            }
+        }
+    }
+
+    pub fn scroll_down_half_page(&mut self) {
+        if !self.is_loaded() || self.total_lines() == 0 {
+            return;
+        }
+        let half = self.content_height as usize / 2;
+        if let Some(index) = self.get_visual_height_index() {
+            let total_vr = index.total_visual_rows();
+            if total_vr == 0 {
+                return;
+            }
+            let cur_vr = self.cursor_visual_row();
+            let target_vr = cur_vr.saturating_add(half as u64).min(total_vr - 1);
+            let (new_line, new_sub) = index.visual_row_to_logical_row_with_offset(target_vr);
+            self.cursor_line = new_line;
+            self.cursor_sub_offset = new_sub;
+            self.ensure_cursor_visible();
+        } else {
+            let width = self.get_content_width();
+            if width > 0 && half > 0 {
+                let (new_line, new_sub) =
+                    self.advance_visual_pos(self.cursor_line, self.cursor_sub_offset, half, width);
+                self.cursor_line = new_line;
+                self.cursor_sub_offset = new_sub;
+                self.ensure_cursor_visible_no_vhi();
+            }
+        }
+    }
+
+    pub fn scroll_up_half_page(&mut self) {
+        if !self.is_loaded() || self.total_lines() == 0 {
+            return;
+        }
+        let half = self.content_height as usize / 2;
+        if let Some(index) = self.get_visual_height_index() {
+            let total_vr = index.total_visual_rows();
+            if total_vr == 0 {
+                return;
+            }
+            let cur_vr = self.cursor_visual_row();
+            let target_vr = cur_vr.saturating_sub(half as u64);
+            let (new_line, new_sub) = index.visual_row_to_logical_row_with_offset(target_vr);
+            self.cursor_line = new_line;
+            self.cursor_sub_offset = new_sub;
+            self.ensure_cursor_visible();
+        } else {
+            let width = self.get_content_width();
+            if width > 0 && half > 0 {
+                let (new_line, new_sub) =
+                    self.retreat_visual_pos(self.cursor_line, self.cursor_sub_offset, half, width);
+                self.cursor_line = new_line;
+                self.cursor_sub_offset = new_sub;
+                self.ensure_cursor_visible_no_vhi();
+            }
+        }
+    }
+
+    pub fn scroll_down_page(&mut self) {
+        if !self.is_loaded() || self.total_lines() == 0 {
+            return;
+        }
+        let page = self.content_height as usize;
+        if let Some(index) = self.get_visual_height_index() {
+            let total_vr = index.total_visual_rows();
+            if total_vr == 0 {
+                return;
+            }
+            let cur_vr = self.cursor_visual_row();
+            let target_vr = cur_vr.saturating_add(page as u64).min(total_vr - 1);
+            let (new_line, new_sub) = index.visual_row_to_logical_row_with_offset(target_vr);
+            self.cursor_line = new_line;
+            self.cursor_sub_offset = new_sub;
+            self.ensure_cursor_visible();
+        } else {
+            let width = self.get_content_width();
+            if width > 0 && page > 0 {
+                let (new_line, new_sub) =
+                    self.advance_visual_pos(self.cursor_line, self.cursor_sub_offset, page, width);
+                self.cursor_line = new_line;
+                self.cursor_sub_offset = new_sub;
+                self.ensure_cursor_visible_no_vhi();
+            }
+        }
+    }
+
+    pub fn scroll_up_page(&mut self) {
+        if !self.is_loaded() || self.total_lines() == 0 {
+            return;
+        }
+        let page = self.content_height as usize;
+        if let Some(index) = self.get_visual_height_index() {
+            let total_vr = index.total_visual_rows();
+            if total_vr == 0 {
+                return;
+            }
+            let cur_vr = self.cursor_visual_row();
+            let target_vr = cur_vr.saturating_sub(page as u64);
+            let (new_line, new_sub) = index.visual_row_to_logical_row_with_offset(target_vr);
+            self.cursor_line = new_line;
+            self.cursor_sub_offset = new_sub;
+            self.ensure_cursor_visible();
+        } else {
+            let width = self.get_content_width();
+            if width > 0 && page > 0 {
+                let (new_line, new_sub) =
+                    self.retreat_visual_pos(self.cursor_line, self.cursor_sub_offset, page, width);
+                self.cursor_line = new_line;
+                self.cursor_sub_offset = new_sub;
+                self.ensure_cursor_visible_no_vhi();
+            }
+        }
+    }
+
+    pub fn scroll_to_top(&mut self) {
+        if !self.is_loaded() || self.total_lines() == 0 {
+            return;
+        }
+        self.cursor_line = 0;
+        self.v_offset = 0;
+        self.v_sub_offset = 0;
+        self.cursor_sub_offset = 0;
+    }
+
+    pub fn scroll_to_bottom(&mut self) {
+        if !self.is_loaded() || self.total_lines() == 0 {
+            return;
+        }
+        self.cursor_line = self.total_lines().saturating_sub(1);
+        if let Some(idx) = self.get_visual_height_index() {
+            self.cursor_sub_offset = idx
+                .visual_height_of_line(self.cursor_line)
+                .saturating_sub(1);
+            self.v_sub_offset = 0;
+            self.ensure_cursor_visible();
+            self.clamp_v_offset();
+        } else {
+            let width = self.get_content_width();
+            self.cursor_sub_offset = if width > 0 {
+                self.line_visual_height(self.cursor_line, width)
+                    .saturating_sub(1)
+            } else {
+                0
+            };
+            self.ensure_cursor_visible_no_vhi();
+        }
+    }
+
+    pub(super) fn ensure_cursor_visible(&mut self) {
+        if !self.is_loaded() || self.total_lines() == 0 || self.content_height == 0 {
+            return;
+        }
+        if self.is_loading() {
+            return;
+        }
+        let cursor_visual = self.cursor_visual_row() as usize;
+
+        let content_h = self.content_height as usize;
+
+        if cursor_visual < self.v_offset {
+            self.v_offset = cursor_visual;
+        } else if cursor_visual >= self.v_offset.saturating_add(content_h) {
+            self.v_offset = cursor_visual.saturating_sub(content_h).saturating_add(1);
+        }
+        self.clamp_v_offset();
+    }
+
+    pub(super) fn clamp_v_offset(&mut self) {
+        let max_offset = self
+            .total_visual_rows()
+            .saturating_sub(self.content_height as usize);
+        self.v_offset = self.v_offset.min(max_offset);
+    }
+
+    fn cursor_visual_row(&self) -> u64 {
+        if let Some(index) = self.get_visual_height_index() {
+            let first = index.cursor_to_first_visual_row(self.cursor_line);
+            let height = index.visual_height_of_line(self.cursor_line);
+            let max_sub = height.saturating_sub(1);
+            let sub = self.cursor_sub_offset.min(max_sub) as u64;
+            first + sub
+        } else {
+            self.cursor_line as u64
+        }
+    }
+}

crates/tui/src/app/settings.rs

@@ -0,0 +1,98 @@
+use crossterm::event::{KeyCode, KeyEvent};
+
+use crate::color::AVAILABLE_COLORS;
+
+use super::{App, AppMode};
+
+impl App {
+    pub(super) fn handle_settings_key(&mut self, key: KeyEvent) {
+        match key.code {
+            KeyCode::Esc | KeyCode::Char('q') => {
+                self.settings_error = None;
+                self.mode = AppMode::Normal;
+            }
+            KeyCode::Enter => {
+                let draft = self.settings_draft.clone();
+                match draft.save() {
+                    Ok(()) => {
+                        self.color_config = draft;
+                        self.settings_error = None;
+                        self.mode = AppMode::Normal;
+                    }
+                    Err(e) => {
+                        self.settings_error = Some(format!("Failed to save settings: {e}"));
+                    }
+                }
+            }
+            KeyCode::Char('j') | KeyCode::Down => {
+                if self.settings_cursor < 5 {
+                    self.settings_cursor += 1;
+                }
+            }
+            KeyCode::Char('k') | KeyCode::Up => {
+                self.settings_cursor = self.settings_cursor.saturating_sub(1);
+            }
+            KeyCode::Left => {
+                self.cycle_color(self.settings_cursor, false);
+            }
+            KeyCode::Right => {
+                self.cycle_color(self.settings_cursor, true);
+            }
+            KeyCode::Char(c) if ('1'..='8').contains(&c) => {
+                let idx = (c as usize) - ('1' as usize);
+                self.set_color(self.settings_cursor, AVAILABLE_COLORS[idx]);
+            }
+            _ => {}
+        }
+    }
+
+    fn cycle_color(&mut self, level_idx: usize, forward: bool) {
+        let current = self.get_settings_color(level_idx).to_string();
+        let colors = AVAILABLE_COLORS;
+        let pos = colors.iter().position(|&c| c == current);
+        let new_pos = match pos {
+            Some(p) => {
+                if forward {
+                    (p + 1) % colors.len()
+                } else if p == 0 {
+                    colors.len() - 1
+                } else {
+                    p - 1
+                }
+            }
+            None => {
+                if forward {
+                    0
+                } else {
+                    colors.len() - 1
+                }
+            }
+        };
+        self.set_color(level_idx, colors[new_pos]);
+    }
+
+    fn get_settings_color(&self, level_idx: usize) -> &str {
+        match level_idx {
+            0 => &self.settings_draft.error,
+            1 => &self.settings_draft.warn,
+            2 => &self.settings_draft.info,
+            3 => &self.settings_draft.debug,
+            4 => &self.settings_draft.trace,
+            5 => &self.settings_draft.unknown,
+            _ => "white",
+        }
+    }
+
+    fn set_color(&mut self, level_idx: usize, color_name: &str) {
+        self.settings_error = None;
+        match level_idx {
+            0 => self.settings_draft.error = color_name.to_string(),
+            1 => self.settings_draft.warn = color_name.to_string(),
+            2 => self.settings_draft.info = color_name.to_string(),
+            3 => self.settings_draft.debug = color_name.to_string(),
+            4 => self.settings_draft.trace = color_name.to_string(),
+            5 => self.settings_draft.unknown = color_name.to_string(),
+            _ => {}
+        }
+    }
+}

crates/tui/src/app/viewport.rs

@@ -0,0 +1,222 @@
+use std::borrow::Cow;
+
+use log_viewer_core::io::wrap::{MAX_WRAP_INPUT_LEN, format_json_line, wrap_line_chars};
+
+use super::viewport_cache::{ViewportEntry, truncate_to_columns};
+use super::{App, AppLoadingState};
+
+impl App {
+    /// Compute a single line's viewport entry (wrapped rows + level + height).
+    pub(super) fn compute_line_entry(&self, line: usize, width: usize) -> ViewportEntry {
+        let raw = self.get_line(line).unwrap_or_default();
+
+        // Guard 1: oversized raw input — skip detect_level and JSON formatting
+        // to avoid O(n) parsing overhead on huge lines.
+        if raw.len() > MAX_WRAP_INPUT_LEN {
+            return ViewportEntry {
+                wrapped_rows: vec![truncate_to_columns(&raw, width)],
+                level: None,
+                visual_height: 1,
+            };
+        }
+
+        let level = log_viewer_core::parser::level::detect_level(&raw);
+        let display_text: Cow<'_, str> = if self.json_format {
+            format_json_line(&raw)
+        } else {
+            Cow::Borrowed(raw.as_str())
+        };
+
+        // Guard 2: JSON pretty-printing may expand a line beyond the limit.
+        if display_text.len() > MAX_WRAP_INPUT_LEN {
+            return ViewportEntry {
+                wrapped_rows: vec![truncate_to_columns(&display_text, width)],
+                level,
+                visual_height: 1,
+            };
+        }
+
+        let mut wrapped = Vec::new();
+        for sub_line in display_text.split('\n') {
+            wrapped.extend(wrap_line_chars(sub_line, width));
+        }
+        let visual_height = wrapped.len().max(1);
+        ViewportEntry {
+            wrapped_rows: wrapped,
+            level,
+            visual_height,
+        }
+    }
+
+    /// Compute visual height for a single line without storing it.
+    pub(super) fn compute_visual_height(&self, line: usize, width: usize) -> usize {
+        let raw = self.get_line(line).unwrap_or_default();
+
+        // Guard 1: oversized raw input.
+        if raw.len() > MAX_WRAP_INPUT_LEN {
+            return 1;
+        }
+
+        let display_text: Cow<'_, str> = if self.json_format {
+            format_json_line(&raw)
+        } else {
+            Cow::Borrowed(raw.as_str())
+        };
+
+        if display_text.len() > MAX_WRAP_INPUT_LEN {
+            return 1;
+        }
+
+        let mut height = 0;
+        for sub_line in display_text.split('\n') {
+            height += wrap_line_chars(sub_line, width).len();
+        }
+        height.max(1)
+    }
+
+    /// Find (logical_line, offset_in_line) for a given visual row offset.
+    fn find_logical_line_at_visual_row(&self, visual_row: usize, _width: usize) -> (usize, usize) {
+        if let Some(index) = self.get_visual_height_index() {
+            return index.visual_row_to_logical_row_with_offset(visual_row as u64);
+        }
+        (visual_row.min(self.total_lines().saturating_sub(1)), 0)
+    }
+
+    /// Ensure the viewport cache covers the visible range.
+    /// Returns (start_logical, offset_in_line) for rendering.
+    pub(crate) fn ensure_viewport_cache(&mut self, width: usize) -> (usize, usize) {
+        let viewport_height = self.content_height as usize;
+
+        if !self.is_loaded() || width == 0 || viewport_height == 0 {
+            return (0, 0);
+        }
+
+        let params_changed = self.viewport_cache.needs_recompute(width, self.json_format);
+
+        if params_changed {
+            self.viewport_cache.invalidate();
+            self.viewport_cache.width = width;
+            self.viewport_cache.set_json_format(self.json_format);
+            self.ensure_visual_height_index(width);
+
+            if self.get_visual_height_index().is_some() {
+                self.ensure_cursor_visible();
+            } else {
+                self.ensure_cursor_visible_no_vhi();
+            }
+        }
+
+        // Find start logical line from v_offset
+        let (start_logical, offset_in_line) = if self.is_loading() {
+            (
+                self.v_offset.min(self.total_lines().saturating_sub(1)),
+                self.v_sub_offset,
+            )
+        } else {
+            self.find_logical_line_at_visual_row(self.v_offset, width)
+        };
+
+        // Compute viewport entries
+        self.viewport_cache.entries.clear();
+        self.viewport_cache.logical_start = start_logical;
+
+        let total = self.total_lines();
+        let mut rows_remaining = viewport_height + offset_in_line;
+
+        for line_idx in start_logical..total {
+            if rows_remaining == 0 {
+                break;
+            }
+            let entry = self.compute_line_entry(line_idx, width);
+            rows_remaining = rows_remaining.saturating_sub(entry.visual_height);
+            self.viewport_cache.entries.push(entry);
+        }
+
+        (start_logical, offset_in_line)
+    }
+
+    /// Compute total visual rows (cached, lazily evaluated).
+    /// Returns `total_lines` for sampling mode (1:1 mapping).
+    pub(super) fn total_visual_rows(&mut self) -> usize {
+        if self.is_loading() {
+            return self.total_lines();
+        }
+        if let Some(index) = self.get_visual_height_index() {
+            return index.total_visual_rows() as usize;
+        }
+        self.total_lines()
+    }
+
+    pub(super) fn cursor_to_first_visual_row(&self, line: usize) -> usize {
+        if self.is_loading() {
+            return line;
+        }
+        if let Some(index) = self.get_visual_height_index() {
+            return index.cursor_to_first_visual_row(line) as usize;
+        }
+        line
+    }
+
+    pub(super) fn visual_row_to_logical_row(&self, visual_row: usize) -> usize {
+        if self.is_loading() {
+            return visual_row.min(self.total_lines().saturating_sub(1));
+        }
+        if let Some(index) = self.get_visual_height_index() {
+            return index.visual_row_to_logical_row(visual_row as u64);
+        }
+        visual_row.min(self.total_lines().saturating_sub(1))
+    }
+
+    /// Compute the rebased offset pair (logical_line, sub_row) from the
+    /// current visual-row `v_offset`.  Returns `(v_offset, v_sub_offset)`
+    /// suitable for the no-VHI fallback scrolling path.
+    ///
+    /// MUST be called before borrowing `&mut self.loading_state` for
+    /// `invalidate_visual_height_index`, because it reads VHI through
+    /// `&self`.
+    pub(super) fn rebase_offset_for_invalidate(&self) -> (usize, usize) {
+        if self.get_visual_height_index().is_some() {
+            let top_visual = self.v_offset;
+            let top_line = self.visual_row_to_logical_row(top_visual);
+            let line_first_visual = self.cursor_to_first_visual_row(top_line);
+            let sub = top_visual.saturating_sub(line_first_visual);
+            (top_line, sub)
+        } else {
+            (self.v_offset, self.v_sub_offset)
+        }
+    }
+
+    pub(super) fn ensure_visual_height_index(&mut self, width: usize) {
+        let needs_rebuild = match self.get_visual_height_index() {
+            Some(idx) => !idx.is_valid_for(self.json_format, width),
+            None => true,
+        };
+
+        if needs_rebuild {
+            let (new_offset, new_sub) = self.rebase_offset_for_invalidate();
+            if let AppLoadingState::Ready { reader } = &mut self.loading_state {
+                reader.invalidate_visual_height_index();
+                reader.start_visual_height_rebuild(width, self.json_format);
+            }
+            self.v_offset = new_offset;
+            self.v_sub_offset = new_sub;
+            self.clamp_cursor_sub_offset_no_vhi();
+        }
+    }
+
+    pub(super) fn toggle_json_format(&mut self) {
+        self.json_format = !self.json_format;
+        self.viewport_cache.invalidate();
+        let width = self.viewport_cache.width;
+        let (new_offset, new_sub) = self.rebase_offset_for_invalidate();
+        if let AppLoadingState::Ready { reader } = &mut self.loading_state {
+            reader.invalidate_visual_height_index();
+            if width > 0 {
+                reader.start_visual_height_rebuild(width, self.json_format);
+            }
+        }
+        self.v_offset = new_offset;
+        self.v_sub_offset = new_sub;
+        self.clamp_cursor_sub_offset_no_vhi();
+    }
+}

crates/tui/src/app/viewport_cache.rs

@@ -0,0 +1,99 @@
+use log_viewer_core::types::LogLevel;
+use unicode_width::UnicodeWidthChar;
+
+pub(super) struct ViewportEntry {
+    pub(super) wrapped_rows: Vec<String>,
+    pub(super) level: Option<LogLevel>,
+    pub(super) visual_height: usize,
+}
+
+pub(super) struct ViewportCache {
+    pub(super) entries: Vec<ViewportEntry>,
+    pub(super) logical_start: usize,
+    pub(super) width: usize,
+    json_format: bool,
+    pub(super) cached_total_visual_rows: Option<usize>,
+}
+
+impl ViewportCache {
+    pub(super) fn new() -> Self {
+        Self {
+            entries: Vec::new(),
+            logical_start: 0,
+            width: 0,
+            json_format: false,
+            cached_total_visual_rows: None,
+        }
+    }
+
+    pub(super) fn invalidate(&mut self) {
+        self.entries.clear();
+        self.logical_start = 0;
+        self.width = 0;
+        self.cached_total_visual_rows = None;
+    }
+
+    pub(super) fn needs_recompute(&self, width: usize, json_format: bool) -> bool {
+        self.width != width || self.json_format != json_format
+    }
+
+    pub(super) fn set_json_format(&mut self, json_format: bool) {
+        self.json_format = json_format;
+    }
+
+    #[allow(dead_code)]
+    pub(super) fn get_entry(&self, logical_line: usize) -> Option<&ViewportEntry> {
+        if logical_line >= self.logical_start {
+            let idx = logical_line - self.logical_start;
+            self.entries.get(idx)
+        } else {
+            None
+        }
+    }
+}
+
+pub(super) const TRUNCATE_TAB_WIDTH: usize = 4;
+
+pub(super) fn gutter_width_for(total_lines: usize, is_loading: bool) -> usize {
+    if total_lines == 0 {
+        return 0;
+    }
+    let line_num_width = total_lines.to_string().len();
+    let loading_extra = if is_loading { 1 } else { 0 };
+    line_num_width + loading_extra + 1 + 1
+}
+
+pub(super) fn truncate_to_columns(s: &str, max_cols: usize) -> String {
+    if max_cols == 0 || s.is_empty() {
+        return String::new();
+    }
+
+    let mut out = String::new();
+    let mut col = 0;
+
+    for ch in s.chars() {
+        if ch == '\t' {
+            let tab_stop = TRUNCATE_TAB_WIDTH - (col % TRUNCATE_TAB_WIDTH);
+            if col + tab_stop > max_cols {
+                break;
+            }
+            for _ in 0..tab_stop {
+                out.push(' ');
+            }
+            col += tab_stop;
+        } else {
+            let w = if ch.is_control() {
+                0
+            } else {
+                ch.width().unwrap_or(0)
+            };
+            if col + w > max_cols {
+                break;
+            }
+            out.push(ch);
+            col += w;
+        }
+    }
+
+    out
+}

crates/tui/src/app/watcher.rs

@@ -0,0 +1,162 @@
+use log_viewer_core::io::file_reader::AppendStatus;
+use log_viewer_core::io::progressive_reader::{ReaderState, compute_line_visual_height};
+use log_viewer_core::watcher::file_watcher::FileEvent;
+
+use super::viewport_cache::gutter_width_for;
+use super::{App, AppLoadingState};
+
+impl App {
+    pub fn poll_file_watcher(&mut self) {
+        let events: Vec<FileEvent> = match &mut self.file_watcher {
+            Some(w) => std::iter::from_fn(|| w.try_recv()).collect(),
+            None => return,
+        };
+
+        for event in events {
+            match event {
+                FileEvent::Appended { new_size: _ } => {
+                    self.handle_file_appended();
+                }
+                FileEvent::Truncated { new_size: _ } => {
+                    self.handle_file_truncated();
+                }
+                FileEvent::Rotated { new_inode: _ } => {
+                    // Don't auto-switch; old content preserved.
+                    // User can reload manually if desired.
+                }
+                FileEvent::Removed => {
+                    self.loading_state = AppLoadingState::Error("File has been deleted".into());
+                }
+                FileEvent::WatcherError { message: _ } => {}
+            }
+        }
+    }
+
+    pub(super) fn handle_file_appended(&mut self) {
+        let rebased = self.rebase_offset_for_invalidate();
+        match &mut self.loading_state {
+            AppLoadingState::Ready { reader } => {
+                let old_reader_line_count = reader.line_count();
+                let status = reader.update_for_append();
+                let width = {
+                    let total = if self.content_width > 0 {
+                        self.content_width as usize
+                    } else {
+                        80
+                    };
+                    total.saturating_sub(gutter_width_for(reader.line_count(), false))
+                };
+                match status {
+                    Ok(AppendStatus::Appended(_new_lines)) => {
+                        let _ = reader.save_cache();
+
+                        let (old_line_count, can_extend) = {
+                            match &reader.state {
+                                ReaderState::Ready {
+                                    visual_height_index: Some(idx),
+                                    ..
+                                } => (idx.line_count(), idx.is_valid_for(self.json_format, width)),
+                                _ => (0, false),
+                            }
+                        };
+                        let new_line_count = reader.line_count();
+
+                        if can_extend && old_line_count == old_reader_line_count {
+                            if let ReaderState::Ready {
+                                visual_height_index: Some(index),
+                                reader: fr,
+                            } = &mut reader.state
+                            {
+                                if old_line_count > 0 {
+                                    let last_old_line_text =
+                                        fr.get_line(old_line_count - 1).unwrap_or("");
+                                    let new_h = compute_line_visual_height(
+                                        last_old_line_text,
+                                        width,
+                                        self.json_format,
+                                    );
+                                    index.replace_last_line_height(new_h);
+                                }
+                                let mut new_heights = Vec::with_capacity(
+                                    new_line_count.saturating_sub(old_line_count),
+                                );
+                                for i in old_line_count..new_line_count {
+                                    let line_text = fr.get_line(i).unwrap_or("");
+                                    new_heights.push(compute_line_visual_height(
+                                        line_text,
+                                        width,
+                                        self.json_format,
+                                    ));
+                                }
+                                index.extend_from_heights(&new_heights);
+                            }
+                        } else {
+                            let (new_offset, new_sub) = rebased;
+                            self.v_offset = new_offset;
+                            self.v_sub_offset = new_sub;
+                            self.cursor_sub_offset = 0;
+                            reader.invalidate_visual_height_index();
+                            reader.start_visual_height_rebuild(width, self.json_format);
+                        }
+
+                        self.viewport_cache.invalidate();
+                    }
+                    Ok(AppendStatus::Reloaded) => {
+                        let _ = reader.save_cache();
+                        let (new_offset, _new_sub) = rebased;
+                        reader.invalidate_visual_height_index();
+                        reader.start_visual_height_rebuild(width, self.json_format);
+                        self.cursor_line =
+                            self.cursor_line.min(self.total_lines().saturating_sub(1));
+                        self.v_offset = new_offset;
+                        self.v_sub_offset = 0;
+                        self.cursor_sub_offset = 0;
+                        self.viewport_cache.invalidate();
+                        self.clamp_v_offset();
+                    }
+                    Ok(AppendStatus::Unchanged) | Err(_) => {}
+                }
+            }
+            AppLoadingState::Loading { .. } => {
+                self.reload_after_loading = true;
+            }
+            _ => {}
+        }
+    }
+
+    pub(super) fn reload_ready_reader(&mut self) {
+        let (new_offset, _new_sub) = self.rebase_offset_for_invalidate();
+        if let AppLoadingState::Ready { reader } = &mut self.loading_state {
+            let _ = reader.reload();
+            let width = {
+                let total = if self.content_width > 0 {
+                    self.content_width as usize
+                } else {
+                    80
+                };
+                total.saturating_sub(gutter_width_for(reader.line_count(), false))
+            };
+            let _ = reader.save_cache();
+            reader.invalidate_visual_height_index();
+            reader.start_visual_height_rebuild(width, self.json_format);
+            self.cursor_line = self.cursor_line.min(self.total_lines().saturating_sub(1));
+            self.v_offset = new_offset;
+            self.v_sub_offset = 0;
+            self.clamp_cursor_sub_offset_no_vhi();
+            self.viewport_cache.invalidate();
+            self.clamp_v_offset();
+        }
+    }
+
+    pub(super) fn handle_file_truncated(&mut self) {
+        match &mut self.loading_state {
+            AppLoadingState::Ready { .. } => {
+                self.reload_ready_reader();
+            }
+            AppLoadingState::Loading { .. } => {
+                self.reload_after_loading = true;
+            }
+            _ => {}
+        }
+    }
+}

crates/tui/src/main.rs

@@ -1,3 +1,6 @@
+use std::time::Duration;
+
+use anyhow::Context;
 use clap::Parser;
 
 mod app;
@@ -11,44 +14,93 @@ struct Cli {
     files: Vec<String>,
 }
 
+// ── RAII terminal guard ────────────────────────────────────────────
+//
+// Holds the ratatui Terminal and guarantees terminal state restoration
+// (disable raw mode, leave alternate screen, show cursor) on **every**
+// exit path: normal return, `?` error propagation, and panic unwind.
+//
+// IMPORTANT: `Drop` does **not** run on `std::process::exit` or
+// `panic = "abort"`.  Therefore we must never call `process::exit`
+// inside the guarded scope — use `?` to return `Err` instead.
+
+type Backend = ratatui::backend::CrosstermBackend<std::io::Stdout>;
+
+struct TerminalGuard {
+    terminal: ratatui::Terminal<Backend>,
+}
+
+impl TerminalGuard {
+    /// Enable raw mode, enter alternate screen, and create the terminal.
+    /// On partial failure, rolls back any steps that already succeeded.
+    fn enter() -> anyhow::Result<Self> {
+        crossterm::terminal::enable_raw_mode()
+            .map_err(|e| anyhow::anyhow!("enable_raw_mode: {e}"))?;
+
+        let mut stdout = std::io::stdout();
+        if let Err(e) = crossterm::execute!(stdout, crossterm::terminal::EnterAlternateScreen) {
+            let _ = crossterm::terminal::disable_raw_mode();
+            return Err(anyhow::anyhow!("EnterAlternateScreen: {e}"));
+        }
+
+        let backend = ratatui::backend::CrosstermBackend::new(stdout);
+        match ratatui::Terminal::new(backend) {
+            Ok(terminal) => Ok(Self { terminal }),
+            Err(e) => {
+                // Roll back: leave alternate screen + disable raw mode
+                let _ = crossterm::execute!(
+                    std::io::stdout(),
+                    crossterm::terminal::LeaveAlternateScreen
+                );
+                let _ = crossterm::terminal::disable_raw_mode();
+                Err(anyhow::anyhow!("Terminal::new: {e}"))
+            }
+        }
+    }
+
+    fn terminal(&mut self) -> &mut ratatui::Terminal<Backend> {
+        &mut self.terminal
+    }
+}
+
+impl Drop for TerminalGuard {
+    fn drop(&mut self) {
+        // Best-effort cleanup; suppress errors (Drop must not panic).
+        let _ = crossterm::terminal::disable_raw_mode();
+        let _ = crossterm::execute!(
+            self.terminal.backend_mut(),
+            crossterm::terminal::LeaveAlternateScreen
+        );
+        let _ = self.terminal.show_cursor();
+    }
+}
+
+// ── main ───────────────────────────────────────────────────────────
+
 fn main() -> anyhow::Result<()> {
     let cli = Cli::parse();
 
-    crossterm::terminal::enable_raw_mode()?;
+    let mut guard = TerminalGuard::enter()?;
-    let mut stdout = std::io::stdout();
-    crossterm::execute!(stdout, crossterm::terminal::EnterAlternateScreen)?;
-    let backend = ratatui::backend::CrosstermBackend::new(stdout);
-    let mut terminal = ratatui::Terminal::new(backend)?;
-
     let mut app = app::App::new();
     app.color_config = log_viewer_core::config::ColorConfig::load();
 
-    if let Some(file) = cli.files.first()
+    if let Some(file) = cli.files.first() {
-        && let Err(e) = app.load_file(file)
+        app.load_file(file)
-    {
+            .with_context(|| format!("loading file {file}"))?;
-        eprintln!("Error loading file: {e}");
-        std::process::exit(1);
     }
 
     while !app.should_quit {
         app.poll_background_indexer();
         app.poll_file_watcher();
-        terminal.draw(|frame| ui::render(frame, &mut app))?;
+        guard.terminal().draw(|frame| ui::render(frame, &mut app))?;
-        if crossterm::event::poll(std::time::Duration::from_millis(100))? {
+        if crossterm::event::poll(Duration::from_millis(100))? {
             match crossterm::event::read()? {
                 crossterm::event::Event::Key(key) => app.handle_key(key),
-                crossterm::event::Event::Resize(_w, _h) => {}
+                crossterm::event::Event::Resize(_, _) => {}
                 _ => {}
             }
         }
     }
 
-    crossterm::terminal::disable_raw_mode()?;
-    crossterm::execute!(
-        terminal.backend_mut(),
-        crossterm::terminal::LeaveAlternateScreen
-    )?;
-    terminal.show_cursor()?;
-
     Ok(())
 }

crates/tui/src/ui.rs

@@ -6,6 +6,7 @@ use crate::color::level_fg;
 use log_viewer_core::config::ColorConfig;
 use log_viewer_core::types::LogLevel;
 
+#[allow(dead_code)]
 pub(crate) fn build_line_spans(
     gutter_text: String,
     content_text: String,
@@ -28,6 +29,10 @@ pub(crate) fn build_line_spans(
     ])
 }
 
+fn is_cursor_visual_row(app: &App, logical_line: usize, visual_row: usize) -> bool {
+    logical_line == app.cursor_line() && visual_row == app.cursor_sub_offset()
+}
+
 pub fn render(frame: &mut ratatui::Frame, app: &mut App) {
     use ratatui::layout::{Constraint, Layout};
     use ratatui::widgets::Paragraph;
@@ -40,7 +45,7 @@ pub fn render(frame: &mut ratatui::Frame, app: &mut App) {
     .split(frame.area());
 
     // ── Title bar ──────────────────────────────────────────────────
-    let title_text = if app.mode == AppMode::Settings {
+    let title_text = if app.mode() == AppMode::Settings {
         " Color Settings".to_string()
     } else if app.is_loading() {
         let name = app.file_name().unwrap_or("unknown");
@@ -51,7 +56,7 @@ pub fn render(frame: &mut ratatui::Frame, app: &mut App) {
         let cursor_display = if app.total_lines() == 0 {
             0
         } else {
-            app.cursor_line + 1
+            app.cursor_line() + 1
         };
         format!(" {} [{}/{}]", name, cursor_display, app.total_lines())
     } else {
@@ -63,7 +68,7 @@ pub fn render(frame: &mut ratatui::Frame, app: &mut App) {
     );
 
     // ── Content area ───────────────────────────────────────────────
-    if app.mode == AppMode::Settings {
+    if app.mode() == AppMode::Settings {
         render_settings(frame, app, outer[1]);
     } else if app.is_error() {
         let msg = app.error_message().unwrap_or_default();
@@ -88,9 +93,21 @@ pub fn render(frame: &mut ratatui::Frame, app: &mut App) {
     }
 
     // ── Status bar ─────────────────────────────────────────────────
-    let status_text = if app.mode == AppMode::Settings {
+    if app.mode() == AppMode::Settings {
-        " j/k:navigate  ←/→:change  1-8:jump  Enter:save  Esc:cancel"
+        if let Some(err) = app.settings_error() {
-    } else if app.is_error() {
+            frame.render_widget(
+                Paragraph::new(err).style(Style::default().fg(Color::Red)),
+                outer[2],
+            );
+        } else {
+            frame.render_widget(
+                Paragraph::new(" j/k:navigate  ←/→:change  1-8:jump  Enter:save  Esc:cancel"),
+                outer[2],
+            );
+        }
+        return;
+    }
+    let status_text = if app.is_error() {
         " Press q to quit"
     } else if app.is_loading() {
         let pct = app.loading_progress().map_or(0, |p| p as usize);
@@ -110,8 +127,11 @@ pub fn render(frame: &mut ratatui::Frame, app: &mut App) {
     } else if app.is_loaded() {
         let name = app.file_name().unwrap_or("unknown");
         let total = app.total_lines();
-        let cursor_display = if total == 0 { 0 } else { app.cursor_line + 1 };
+        let cursor_display = if total == 0 { 0 } else { app.cursor_line() + 1 };
-        let status = format!(" {} [{}/{}] | j/k:scroll d/u:half-page f/b:page G/gg:jump Tab:format S:settings q:quit", name, cursor_display, total);
+        let status = format!(
+            " {} [{}/{}] | j/k:scroll d/u:half-page f/b:page G/gg:jump Tab:format S:settings q:quit",
+            name, cursor_display, total
+        );
         frame.render_widget(Paragraph::new(status), outer[2]);
         return;
     } else {
@@ -130,26 +150,31 @@ pub fn render_settings(frame: &mut ratatui::Frame, app: &mut App, area: ratatui:
 
     let popup_w = ((area.width as u32 * 4 / 5).max(40)).min(area.width as u32) as u16;
     let popup_h = ((area.height as u32 * 4 / 5).max(14)).min(area.height as u32) as u16;
-    let popup_x = area.width.saturating_sub(popup_w) / 2;
+    let popup_x = area
-    let popup_y = area.height.saturating_sub(popup_h) / 2;
+        .x
+        .saturating_add(area.width.saturating_sub(popup_w) / 2);
+    let popup_y = area
+        .y
+        .saturating_add(area.height.saturating_sub(popup_h) / 2);
     let popup = ratatui::layout::Rect::new(popup_x, popup_y, popup_w, popup_h);
 
     let block = Block::new().borders(Borders::ALL).title(" Color Settings ");
     let inner = block.inner(popup);
     frame.render_widget(block, popup);
 
+    let settings_draft = app.settings_draft();
     let levels = [
-        ("ERROR", &app.settings_draft.error),
+        ("ERROR", &settings_draft.error),
-        ("WARN", &app.settings_draft.warn),
+        ("WARN", &settings_draft.warn),
-        ("INFO", &app.settings_draft.info),
+        ("INFO", &settings_draft.info),
-        ("DEBUG", &app.settings_draft.debug),
+        ("DEBUG", &settings_draft.debug),
-        ("TRACE", &app.settings_draft.trace),
+        ("TRACE", &settings_draft.trace),
-        ("UNKNOWN", &app.settings_draft.unknown),
+        ("UNKNOWN", &settings_draft.unknown),
     ];
 
     let mut lines = Vec::new();
     for (i, (level_name, color_name)) in levels.iter().enumerate() {
-        let is_selected = i == app.settings_cursor;
+        let is_selected = i == app.settings_cursor();
         let cursor_marker = if is_selected { "▶ " } else { "  " };
 
         let preview_color = color_name.parse::<Color>().unwrap_or(Color::White);
@@ -200,8 +225,7 @@ fn render_content(frame: &mut ratatui::Frame, app: &mut App, area: ratatui::layo
     let content_width = area.width as usize;
     let content_height = area.height as usize;
 
-    app.content_height = area.height;
+    app.set_content_area(area.width, area.height);
-    app.content_width = area.width;
 
     let total_lines = app.total_lines();
     let line_num_width = if total_lines > 0 {
@@ -209,14 +233,8 @@ fn render_content(frame: &mut ratatui::Frame, app: &mut App, area: ratatui::layo
     } else {
         0
     };
-
     let is_loading = app.is_loading();
-    let gutter_prefix_extra = if is_loading { 1 } else { 0 };
+    let gutter_width = app.gutter_width();
-    let gutter_width = if total_lines > 0 {
-        line_num_width + gutter_prefix_extra + 1 + 1
-    } else {
-        0
-    };
 
     let actual_content_width = content_width.saturating_sub(gutter_width);
 
@@ -227,7 +245,6 @@ fn render_content(frame: &mut ratatui::Frame, app: &mut App, area: ratatui::layo
     let (start_logical, offset_in_line) = app.ensure_viewport_cache(actual_content_width);
 
     let mut lines: Vec<Line> = Vec::new();
-    let mut current_visual_offset: usize = 0;
     let available_rows = content_height;
 
     let gutter_style = if is_loading {
@@ -238,40 +255,28 @@ fn render_content(frame: &mut ratatui::Frame, app: &mut App, area: ratatui::layo
         Style::default().fg(Color::DarkGray)
     };
 
-    for (entry_idx, entry) in app.viewport_cache.entries.iter().enumerate() {
+    for row in app.viewport_rows(start_logical, offset_in_line, available_rows) {
-        let logical_line = app.viewport_cache.logical_start + entry_idx;
+        let is_cursor = is_cursor_visual_row(app, row.logical_line, row.visual_row);
-        let start_row = if logical_line == start_logical {
+        let level = row.level;
-            offset_in_line
-        } else {
-            0
-        };
-
-        for (visual_row, text) in entry.wrapped_rows.iter().enumerate().skip(start_row) {
-            if current_visual_offset >= available_rows {
-                break;
-            }
-
-            let is_cursor = logical_line == app.cursor_line;
-            let level = entry.level.as_ref();
 
         let bg_color = if is_cursor {
             Color::DarkGray
         } else {
             Color::Reset
         };
-            let level_fg = level_fg(level, &app.color_config).unwrap_or(Color::White);
+        let level_fg = level_fg(level, app.color_config()).unwrap_or(Color::White);
 
-            let gutter_text = if visual_row == 0 {
+        let gutter_text = if row.visual_row == 0 {
             if is_loading {
                 format!(
                     "~{:>width$} \u{2502}",
-                        logical_line + 1,
+                    row.logical_line + 1,
                     width = line_num_width
                 )
             } else {
                 format!(
                     "{:>width$} \u{2502}",
-                        logical_line + 1,
+                    row.logical_line + 1,
                     width = line_num_width
                 )
             }
@@ -289,15 +294,11 @@ fn render_content(frame: &mut ratatui::Frame, app: &mut App, area: ratatui::layo
 
         lines.push(Line::from(vec![
             Span::styled(gutter_text, effective_gutter_style),
-                Span::styled(text.clone(), Style::default().fg(level_fg).bg(bg_color)),
+            Span::styled(
+                row.text.to_string(),
+                Style::default().fg(level_fg).bg(bg_color),
+            ),
         ]));
-
-            current_visual_offset += 1;
-        }
-
-        if current_visual_offset >= available_rows {
-            break;
-        }
     }
 
     while lines.len() < available_rows {
@@ -363,6 +364,16 @@ mod tests {
         assert_eq!(line.spans[1].style.bg, Some(Color::Reset));
     }
 
+    #[test]
+    fn test_cursor_highlight_predicate_uses_cursor_sub_offset() {
+        let mut app = App::new();
+        app.set_cursor_for_test(4, 2);
+
+        assert!(!is_cursor_visual_row(&app, 4, 0));
+        assert!(!is_cursor_visual_row(&app, 3, 2));
+        assert!(is_cursor_visual_row(&app, 4, 2));
+    }
+
     fn render_to_buffer(app: &mut App, width: u16, height: u16) -> ratatui::buffer::Buffer {
         let backend = ratatui::backend::TestBackend::new(width, height);
         let mut terminal = ratatui::Terminal::new(backend).unwrap();
@@ -469,7 +480,8 @@ mod tests {
         let result = std::panic::catch_unwind(|| {
             let data = std::fs::read(&path).unwrap();
             let index = log_viewer_core::io::line_index::LineIndex::from_bytes(&data);
-            let _ = log_viewer_core::io::index_cache::IndexCache::save(&path, &index);
+            let _ =
+                log_viewer_core::io::index_cache::IndexCache::save_with_hash(&path, &index, &data);
 
             let mut app = App::new();
             app.load_file(path.to_str().unwrap()).unwrap();
@@ -493,4 +505,84 @@ mod tests {
         let _ = std::fs::remove_file(&path);
         assert!(result.is_ok());
     }
+
+    // ── Issue #31: Settings popup area offset tests ────────────────
+
+    /// Helper: enter settings mode and render to buffer.
+    fn render_settings_to_buffer(
+        app: &mut App,
+        width: u16,
+        height: u16,
+    ) -> ratatui::buffer::Buffer {
+        app.enter_settings_mode_for_test();
+        render_to_buffer(app, width, height)
+    }
+
+    /// Find the top-left corner of the popup border by scanning for '┌'.
+    fn find_popup_top_left(
+        buf: &ratatui::buffer::Buffer,
+        width: u16,
+        height: u16,
+    ) -> Option<(u16, u16)> {
+        for row in 0..height {
+            for col in 0..width {
+                if buf.cell((col, row)).unwrap().symbol() == "┌" {
+                    return Some((col, row));
+                }
+            }
+        }
+        None
+    }
+
+    #[test]
+    fn test_settings_popup_includes_area_offset() {
+        // In an 80x24 frame with Layout [Length(1), Min(1), Length(1)]:
+        //   outer[0] = title bar  -> y=0
+        //   outer[1] = content    -> y=1, height=22
+        //   outer[2] = status bar -> y=23
+        // The popup is centered within outer[1], so its y must be >= outer[1].y (which is 1).
+        let mut app = App::new();
+        let buf = render_settings_to_buffer(&mut app, 80, 24);
+
+        let (_px, py) =
+            find_popup_top_left(&buf, 80, 24).expect("popup border '┌' should be rendered");
+
+        // outer[1].y == 1; the popup is centered inside a 22-row area,
+        // so popup_y must be at least 1 (not 0).
+        assert!(
+            py >= 1,
+            "popup top row should account for area.y offset, got y={py} (expected >= 1)"
+        );
+    }
+
+    #[test]
+    fn test_settings_popup_horizontal_centering_uses_area_x() {
+        // outer[1].x is 0 for this layout, so this mainly verifies the popup
+        // is centered and not shifted left. A non-zero area.x layout would
+        // need a different layout to trigger, but the formula is the same.
+        let mut app = App::new();
+        let buf = render_settings_to_buffer(&mut app, 80, 24);
+
+        let (px, _py) =
+            find_popup_top_left(&buf, 80, 24).expect("popup border '┌' should be rendered");
+
+        // popup_w = 80*4/5 = 64, centered: (80-64)/2 = 8
+        assert_eq!(
+            px, 8,
+            "popup should start at x=8 (centered 64-wide popup in 80-col area)"
+        );
+    }
+
+    #[test]
+    fn test_settings_popup_small_frame_no_panic() {
+        // Frame smaller than the min popup size (40x14) should not panic.
+        let result = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
+            let mut app = App::new();
+            let _buf = render_settings_to_buffer(&mut app, 30, 10);
+        }));
+        assert!(
+            result.is_ok(),
+            "rendering settings in a small frame should not panic"
+        );
+    }
 }