Third observed instance (2026-07-31 during the #546 branch eval ×1, 2026-08-04 during v0.33.0 pre-release checks ×1 — both times the 15-model/local-rotation eval suites were hammering the rig network; plus the pattern's first sighting the same Friday).
Signature: fails in ~0.4s (vs ~30s healthy run — a connection-refused-class fast-fail), passes on standalone -count=1 re-run every time. Never observed failing on GitHub CI runners.
Likely fix direction: the test binds/depends on local network resources that contend with concurrent eval traffic — either isolate the port/listener, add a bounded retry for connection setup (NOT for the assertion), or gate it behind a rig-load check. Sibling of #509's flake class (redundant timing guards).
Filed during the v0.33.0 release flow; not release-blocking (CI is the authoritative gate and has never shown it).
Third observed instance (2026-07-31 during the #546 branch eval ×1, 2026-08-04 during v0.33.0 pre-release checks ×1 — both times the 15-model/local-rotation eval suites were hammering the rig network; plus the pattern's first sighting the same Friday).
Signature: fails in ~0.4s (vs ~30s healthy run — a connection-refused-class fast-fail), passes on standalone
-count=1re-run every time. Never observed failing on GitHub CI runners.Likely fix direction: the test binds/depends on local network resources that contend with concurrent eval traffic — either isolate the port/listener, add a bounded retry for connection setup (NOT for the assertion), or gate it behind a rig-load check. Sibling of #509's flake class (redundant timing guards).
Filed during the v0.33.0 release flow; not release-blocking (CI is the authoritative gate and has never shown it).