test(gateway): retain raw wire qualification evidence

openspec/changes/truthful-fixed-frame-qualification/design.md

@@ -8,7 +8,7 @@ Because the bounded fixture uses loopback rather than a physical 1 Gbps link, v8
 
 The retained v6 qualification run passed its then-current checks but is superseded because its tight-loop sender contradicted the pinned Apollo schedule. Private Linux runs 123 and 124 remain failed evidence. One local v8 sustained run passed on Darwin, but it is neither Linux proof nor normative Section 7 evidence.
 
-The later Darwin non-sustained pre-CI invocation was not green and was not retried. Its 1440p120 profile delivered the exact 6,250,000 bytes in 120 frames plus all 6,483 source and warm-up shards with zero drops, but measured 46,973.13 kbps over an implied approximately 1.0644383 seconds and failed the 5% throughput gate. Private Linux full verification/artifact retention and the replacement v9 normative run remain open.
+The later Darwin non-sustained pre-CI invocation was not green and was not retried. Its 1440p120 profile delivered the exact 6,250,000 bytes in 120 frames plus all 6,483 source and warm-up shards with zero drops, but measured 46,973.13 kbps over an implied approximately 1.0644383 seconds and failed the 5% throughput gate. Private Linux full verification/artifact retention and the replacement v10 normative run remain open.
 
 Private Linux run 125 at the frozen v8 harness head is retained as failed evidence. Its exact 33-datagram gap between successful fixture writes and production `MediaIngress` equaled the Linux socket's 33 measured kernel UDP drops. The complete-frame queue, fair pacer, QUIC fragmentation, and public decoder were downstream and did not account for the loss.
 
@@ -22,6 +22,12 @@ Qualification v9 observes every raw public QUIC datagram immediately after the i
 
 Private Linux run 128/job 482 was the single push-triggered attempt at exact source `22433e5c45c179e9d487b59e5d80f1dcf3b285ce`. Linux `make verify`, the sustained gate, strict OpenSpec, deterministic artifact generation, upload, and the clean-checkout step passed. Artifact 28's binaries and SPDX matched the frozen hashes and source metadata. The retained log `/private/tmp/versevdi-gitea-run-128-job-482.log` has SHA-256 `6f5852dce815ba87a55d61aae34cb2fce17a1a62c5ee50e18c0034259ad0a029`. The workflow requested 30 retention days, but Gitea 1.27 floored the positive request delay and the API recorded `2026-08-09T23:18:18+07:00` through `2026-09-07T23:18:18+07:00`, exactly 2,505,600 seconds or 29 elapsed days. The run was not retried. It is passing Linux execution and artifact-byte evidence but retention-nonconforming, so it neither satisfies the private Linux artifact gate nor authorizes normative Section 7. The local 31-day request compensates for verified platform rounding without changing the acceptance threshold; a separately authorized future run must prove an API interval of at least 2,592,000 seconds.
 
+At exact source `c0e362c0285d267f8af4087d07943822311f60e1`, the first Section 7 process created mode-0750 `gateway-rc10-c0e362c` and stopped before fixture startup because the sandbox denied its required loopback bind. The directory remains empty as environment-boundary evidence. The separately authorized escalated attempt retained `gateway-rc10-c0e362c-a2`, passed the v9 runtime checks, and wrote 16 files with manifest SHA-256 `61ea55140dfe6b37332de03879ac33206dd5d47763d09fb0fc2f65bb1f8e02b4`. Independent review denied normative acceptance: the constrained logical CSV and manifest aggregates did not persist every raw public QUIC datagram receive offset/encoded length, and `fairness.csv.gz` lacked manifest transition offsets. A2 therefore remains runtime-passing but normative-raw-evidence-incomplete, without mutation or relabeling.
+
+Qualification v10 keeps the logical impairment CSV unchanged and adds only `impairment-constrained-1080p60-h264-wire.csv.gz`. One monotonic epoch is captured before the constrained delivery and transition sequence. The bounded CSV interleaves two explicit transition records with every public datagram observation in monotonic order using `record_type,reduction_percent,transition_after_ns,received_after_ns,encoded_bytes`; transition-only and delivery-only fields remain empty and are validated as such. Its maximum row count is derived from the existing constrained job bound, the frame-fragment count, and the two configured transitions rather than a captured-run row constant. The manifest binds the file name, SHA-256, compressed bytes, total/delivery/transition row counts, timebase, exact transition offsets, and each capacity summary's recomputation source. Each impairment step is recomputed from every delivery at or after its transition through completion, preserving the original v9 classifier semantics even after the next transition. Fairness remains stage-bounded because `runQualificationFleetStage` records each capacity stage as a separate slice; its summaries retain their 25% and 50% offsets relative to the existing fairness CSV epoch.
+
+The independent parser's caller explicitly selects smoke or normative authority. Normative validation requires exactly 10,000 sent logical units and always enforces the ten-second convergence and 105% rolling-cap gates; retained `sent` data cannot weaken them. Both wire and fairness readers reject compressed input before hashing when it exceeds a writer-derived bound, feed gzip output through a bounded standard-library reader before CSV parsing, and bound fields, offsets, flows, and encoded lengths from the canonical row count, schema, time horizon, and writer values. Aggregate-only v9 data and malformed schema/hash/count/order/transition/length or oversized inputs remain rejected.
+
 The ingress repair follows reviewed behavior rather than copying implementation source:
 
 - Apollo `adc5c5a0bd80831ce495434bb16aee2cd4175fb8`, GPL-3.0, `src/stream.cpp:1463-1474,1573-1627`, supplies the 80%-of-1-Gbps raw-block pacing, 64-KiB/64-packet batch cap, and cross-frame send schedule used by the fixture.
@@ -49,6 +55,8 @@ The native provider therefore requests `2,048 * 1,072 = 2,195,456` bytes with `S
 - Keep video decrypt/FEC single-threaded; only the bounded connected-socket drain is separated so crypto stalls cannot become unexplained kernel loss.
 - Preserve valid scheduling debt after bounded host stalls instead of converting it into provider-queue residence; repay it within the existing fair pacer without a new queue, interface, or configured headroom.
 - Classify constrained-capacity evidence from actual public datagram observations and configured wire capacity; do not infer transport timing from completed logical frames.
+- Persist constrained public-wire observations and transition events in one bounded monotonic-offset CSV, recompute impairment summaries from each transition through completion, and bind both impairment and stage-bounded fairness capacity summaries to their retained raw sources.
+- Select smoke versus normative evidence validation through trusted caller input and bound compressed bytes, decompressed bytes, fields, offsets, flows, and encoded lengths before independent CSV parsing.
 
 ## Risks / Trade-offs