Nightly red since 2026-09-18

[github-actions]

Nightly red: integration failed — https://github.com/GenWave-Org/genwave/actions/runs/35322070451

[github-actions]

Nightly red: integration failed — https://github.com/GenWave-Org/genwave/actions/runs/35430380298

[genwave-radio]

🔍 Root cause — both families trace to one commit

Both red nightlies fail the same integration job (chaos and report are green), with the same tally: Failed 4, Passed 230, Skipped 273, Total 507.

Run	Date	Commit
35322070451	09-18	0ee3c329
35430380298	09-19	2437a011

Last green nightly: 35199554136 (09-17, 5437be31).

The single commit responsible is ca6d432 — "test(specs): T507 delete obsolete and gate facts, report counts manual facts (gh-#777)". It landed after the last green nightly (git merge-base --is-ancestor ca6d432 5437be3 → false).

1️⃣ espeak-ng — 3 facts, Assert.NotNull() Failure: Value is null

tests/GenWave.Host.Tests/Specs/Story324_RespellOracle.cs

T507's goal was "zero gate: facts remain". The skip it deleted was a dynamic one: file static class EspeakNgProbe (a cached Lazy<bool> PATH probe) plus file sealed class RequiresRealEspeakNgAttribute : FactAttribute whose ctor set Skip = "gate: espeak-ng is not on PATH…". The facts became plain [Fact, Trait("Category", "Integration")].

The replacement header comment claims the integration lane runs "inside that image, or with espeak-ng apt-get/brew installed locally". That is not the lane. nightly.yml's integration job runs on a bare ubuntu-latest and its only tool install is ./.github/actions/install-ffmpeg — ffmpeg only. There is not one espeak reference in any workflow file. So the probe was the thing keeping these green, and nothing replaced it.

2️⃣ liquidsoap — 1 fact, exit code 0

tests/GenWave.Host.Tests/Specs/Story129_SafePathLevelMatching.cs:71-102 — TheScriptTypechecksOnPinnedLiquidsoap

liquidsoap --check failed (exit 0): stdout=[] stderr=[Unable to find image 'savonet/liquidsoap:v2.4.4' locally …]

Exit 0 means the type check passed. The assertion also demands stderr.Length == 0, and it measures the stderr of docker run — which writes image-pull progress there on a cold runner. The comment says "silence on stdout/stderr + exit code 0 = pass (F37.1)", but that convention was about liquidsoap's streams, not docker's.

Reproduced locally with a clean image cache:

$ docker rmi busybox:1.36.1-uclibc
$ docker run --rm busybox:1.36.1-uclibc true    # split streams
exit=0  stdout=0 bytes  stderr=325 bytes  ("Unable to find image … locally")

Cold-cache-only, which is why it never showed on a dev box.

🛠 Proposed fixes

1. espeak — add an espeak-ng install step to nightly.yml's integration job (it already has an ffmpeg composite action to pattern-match), or restore a probe-based skip. ⚠️ Note the skip-prefix law from T506/T507 now only permits pending: and manual: — gate: is no longer available, so a restored probe needs a manual: reason.
2. liquidsoap — pre-pull in a separate docker pull savonet/liquidsoap:v2.4.4 step so the --check run's stderr is liquidsoap's alone. (Asserting only on exit code would also work but loses the warning-detection the fact was written for.)

Not touching either file yet — 🙋 Dean's call, and nightly.yml is outside the current branch's scope.

ℹ️ Not this issue

• The 09-16 run failed with Failed: 1 — a different, earlier cause, predating ca6d432.
• gh-Tier 2 red: Story346 real-Postgres verify probe expects db/46's marker that the container never gets #796 (Story346 db/46 marker) is a separate open issue and is not among these 4 failures.