Tests you can trust: no flaky listener races, skips instead of fails without the model, the STOP property test measured from the right moment (866 pass / 0 fail) #504

From a steward bug hunt, 2026-10-21. A triage of every failing test: none were product bugs. All of them were test races, an environment gap, or the wrong measurement. Test files only; product code is unchanged.

• Precision history: the close listener is registered before the STOP (it could miss the close and hang for 5 s). 10 of 10.
• fixtures.ts: HAVE_MODEL, HAVE_PRINTABLES and HAVE_README. Tests that need Simulation/, vision/ or the README skip with a reason instead of failing on partial checkouts.
• STOP property test (a): measured from when the STOP is answered, not pressed. A frame sent in the same breath as the press can legitimately reach the Pi first. The relay's STOP epoch already refuses anything older. New test (d) proves this: 22 of 40 same-tick frames landed after the press, and none landed after the STOP was answered. --rerun-each 50: 400 of 400.
• Lobby ban and mute: every listener is registered before the message that triggers it. Main already has the ban half (0f8a375). This adds the refused-/follow check, which waits for the socket to close before a pong instead of a 150 ms timer.
• Release test: the closing hold can sit 1-2 ticks under the last goal. The tolerance is now 12 ticks, the same as the landing check. 125 of 125.

Suite with the model present: 866 pass, 3 skip (printables and README), 0 fail. On the same base, main has 865 pass and 2 fail.

🤖 Generated with Claude Code