ops(gate): the capture leg — stream to WAV, measured, speech log checked (STORY-445, gh-#777)

[genwave-radio]

PR-5 of the gh-#777 epic (STORY-445, SPEC F178.3, F178.5). The gate now listens to what it launched: a real capture off the stream, measured, with the speech log checked.

🏗️ What

• tools/gate/make_media.sh <outdir> (T489): two 45 s MP3 tones at −12 and −30 LUFS (two-pass loudnorm, verified with ebur128 to ±1 LU) plus two CC0 clips from tools/gate/media/ (≤ 1 MB each, SOURCES.md names creator and URL). The gate's inline synth is gone; it calls this script.
• tools/gate/measure_audio.sh <wav> --target <lufs> (T490): silencedetect (SILENCE_FLOOR −45 dB, SILENCE_SECS 2 s) event count and ebur128 integrated loudness within TOL_LU 5 of the target. One summary line on stdout; exit 0 pass, 1 fail, 2 usage or an input ffmpeg cannot measure.
• The capture leg (T491): after a passed fresh leg, --capture logs in to the api with the scratch's own admin password (grepped from its .env, passed on stdin, never sourced or exported), reads Loudness:TargetLufs from GET /api/settings, records CAPTURE_SECS (default 360) of the stream with ffmpeg -reconnect 1, measures it, and counts station.booth_log rows with a non-null segment_kind in the last 7 minutes. Every number lands in the md and json report; a miss fails the leg naming the assertion (login, target, capture, silence, loudness, measure, booth_log). --capture without --fresh exits 2.
• GateHarness.FakeStation now mirrors the real api: login cookie, 401 without it, settings as an array of {key, value} with a string value.
• Story445_CapturedAudio.cs: all facts green, plus two sad-path scenarios added in review (an off-level capture fails loudness, a silent gap fails silence).

🔌 T492 — the real capture on the dev box (v5.8.3)

tools/gate/stack_gate.sh --tag v5.8.3 --fresh --capture --report <dir>
# exit 0, wall 436 s (fresh leg ~70 s, capture 360 s); docker compose ls | grep gw-gate → empty afterwards

leg	result	detail
fresh	ran/passed	-
upgrade	skipped (--upgrade not given)	-
capture	ran/passed	-
chaos	skipped (--chaos not given)	-

measurement	value
capture secs	360
target	−16 LUFS
silence events	0
integrated	−20.7 LUFS
booth_log rows (speech, last 7 min)	15

The integrated reading sits 4.7 LU under the target, inside the ±5 LU tolerance but close to it. The first minutes of a fresh station are mostly speech and the gate tones at −12/−30 average out under the target, so a longer capture on a fuller library should land nearer −16. Worth an ear on the next LLL rather than a tighter tolerance now.

🔍 Review rounds worth knowing

• T490 round 1: a corrupt WAV exited 183 with nothing on stderr (set -e inside a $(ffmpeg …) capture). Fixed; exit 2 with ffmpeg's tail.
• T491 round 1: if ! v=$(cmd); then s=$? reads the negated pipeline's 0, so a silence or loudness miss reported passed. Fixed to s=0; v=$(cmd) || s=$?, and the two new sad-path facts pin it.

✅ Gate

dotnet test GenWave.sln --filter "Category!=Integration" -- xUnit.MaxParallelThreads=3: all nine projects green (Host 2965 passed / 153 skipped; 0 failed). bash -n and shellcheck clean on all three scripts.

📝 Follow-ups

• The booth_log window is the SPEC's fixed 7 minutes; a CAPTURE_SECS above 420 undercounts (commented in the script).
• launch.sh prunes unused images older than 7 days on success; the gate triggers it. Stop a co-located station with docker compose stop, never down.