fix(gateway): sustain qualified media traversal

openspec/changes/correct-gateway-qualification-evidence/design.md

@@ -23,8 +23,8 @@ The qualification driver already reaches the production Apollo-to-QUIC path, but
 - Assign stable source sequence identifiers and retain per-stage counts so injected loss, provider/FEC drop, queue replacement, QUIC failure, and client miss are disjoint.
 - Reuse the established gateway child-test pattern for the actual gateway server; the Apollo fixture and QUIC client remain in the parent driver. A token-protected loopback test control endpoint starts and stops bounded child-owned recording and returns aggregate stage state.
 - Stream queue, processing, and pacing samples from the production `sendMedia` boundary to child-owned raw evidence. Sample child `RUSAGE_SELF`, Go heap, allocations, and goroutines once per second with independent per-run baselines.
-- Buffer at most 4,096 child-owned timing samples before the gzip writer; drain every sample before recording stops and backpressure on sustained writer overload instead of dropping evidence or compressing synchronously in the media loop.
-- Use the existing reviewed 64-packet qualification bound for native video (about 7.5 ms at 80 Mbps) after public-path counters proved the 16-packet queue replaced 1–5 clean-path units during ordinary scheduler pauses. Keep audio at 16 packets and retain latest-unit replacement at both bounds.
+- Buffer at most 4,096 child-owned timing samples before the gzip writer; encode rows into a reused byte buffer, drain every sample before recording stops, and backpressure on sustained writer overload instead of dropping evidence or formatting/compressing synchronously in the media loop.
+- Bound native video at 256 packets (about 30 ms and less than 0.4 MiB per session at the largest fixture unit) after the sustained public-path regression observed a 141-packet scheduler/GC stall with the 64-packet bound; keep audio and events at 16 and retain latest-unit replacement. Sample heap objects, allocated objects/bytes, and live goroutines through `runtime/metrics` while retaining `RUSAGE_SELF` for CPU.
 - Measure clock overhead as the median elapsed time per read across 1,000 batches of 100 monotonic reads and record that method.
 
 ## Risks / Trade-offs