fix(stats): wire PipeWire drops, expand Display, purge dead residue

Oracle-driven P1 fix plan. Resolves the StatsSnapshot 'computed but never
consumed' debt that was silently zeroing two real diagnostic fields and
leaving a dozen more unreported.

Bug fix (Oracle step 2):
  - state_portal.rs: set_pipewire_dropped(0, 0) and set_queue_depths(0, 0)
    were hardcoded, silently discarding real PipeWire diagnostics. Now wires
    to self.cap.dropped_count() (with pw_dropped_prev delta tracking) and
    self.cap.capture_queue_depth(). The encoded side stays 0 because the
    encoder thread exposes no queue-depth API.

Display expansion (Oracle step 1):
  - stats.rs: StatsSnapshot::Display now reports 12 previously-silent fields
    paired with their existing p95/max counterparts — capture/encoded/sent
    frame counts, elapsed_secs, *_avg_ms gap timing, frame_age_avg_ms,
    per-stage import/sws/encode/total avg_ms, output_frame_bytes_p95. Each
    line of the format string maps to one operational question (cadence,
    drops, queue pressure, latency, bandwidth); layout note added.

Dead residue purge (Oracle steps 5 + 6):
  - stats.rs: removed record_over_budget method + over_budget_count field
    (no caller; total_p95_ms answers the useful question without an
    arbitrary budget threshold).
  - state.rs: removed InFlightSurface::Allocd variant (never constructed)
    and CaptureSource::alloc_frame trait method (prototype leftover; the
    sole impl in cap_wlr_screencopy.rs returned None unconditionally).
  - cap_wlr_screencopy.rs: removed the alloc_frame stub; updated the
    unit-type Frame doc to reference the asynchronicity rationale without
    the deleted method.
  - cap_portal.rs: removed redundant 'let dropped = dropped;' shadowing
    flagged by clippy::redundant_locals (line 849).

Deferred (Oracle step 4 — needs product decision):
  - scale_avg/scale_p95/transfer_avg/transfer_p95/send_wait_p95 fields
    still appear in Display but producers in the live encode path don't
    record them, so they often show misleading zeros. Either add real
    EncState timing for scale/transfer stages, or remove the fields from
    Display until then.

All 79 unit tests + 3 integration tests still pass. clippy: 0 errors.
Warning count: multiple_fields_never_read on StatsSnapshot,
method_never_used on record_over_budget/dropped_count/capture_queue_depth/
alloc_frame, variant_never_constructed on Allocd, redundant_locals on
dropped — all gone.

src/state_portal.rs

@@ -338,8 +338,15 @@ impl StatePortal {
 
         // 每秒输出一次结构化管道统计（仅 --stats 启用时记录日志）
         if self.args.stats && self.stats.should_snapshot() {
-            self.stats.set_pipewire_dropped(0, 0);
-            self.stats.set_queue_depths(0, 0);
+            // Wire PipeWire drop counter (delta-tracked via pw_dropped_prev) and
+            // capture channel depth. Oracle audit 2026-06-28: previously hardcoded
+            // (0, 0), which silently zeroed two real diagnostic fields.
+            let total_dropped = self.cap.dropped_count();
+            self.stats.set_pipewire_dropped(total_dropped, self.pw_dropped_prev);
+            self.pw_dropped_prev = total_dropped;
+            // capture queue depth is real; encoded side has no exposed depth — the
+            // encoder thread publishes timings only, not a frame queue length.
+            self.stats.set_queue_depths(self.cap.capture_queue_depth(), 0);
             if let Some(ref enc_thread) = self.enc_thread {
                 while let Ok(timing) = enc_thread.timing_rx.try_recv() {
                     self.stats.record_encode_thread(