Full sweeps of 2026-09-26 failed and were recorded nowhere; the sweep alert itself failed

[speak-agent]

Summary

The two full sweeps dispatched on main on 2026-09-26, after the pin moved to mcpp 2026.9.26.2 (#470), failed, and no issue recorded either failure. The scheduled sweep of 2026-09-27 passed every member. This issue records the three readings and the two defects that kept the failures unread (the review of 2026-09-28, mcpp-community/mcpp .agents/docs/2026-09-28-ecosystem-review-of-two-days-of-mcpp-and-xlings.md §3.6; task N3 of the design that answers it).

Readings

Run	Event	Commit	Result	Failed
36214011426	workflow_dispatch, 2026-09-26 03:11 UTC	f84c57b	failure	mirror-cn-reachable
36225137392	workflow_dispatch, 2026-09-26 06:54 UTC	366b6ca	failure	workspace (linux default 2/6): pangocairo (FAILED. 1 passed; 1 failed); and full sweep failed — open an issue
36316046499	schedule, 2026-09-27 11:32 UTC	039fb16	success	none
36345994614	workflow_dispatch on mcpp/2026.9.28.1, 2026-09-27 19:53 UTC	e14a3db	cancelled	the sweep for the 2026.9.28.1 pin did not complete

Why nobody read them

1. The alert job failed. full sweep failed — open an issue checks nothing out, so gh issue list could not infer the repository: failed to run git: fatal: not a git repository. The sweep that should have opened an issue opened none. It also did not run for the first failure, because it watched only the workspace job, not mirror-cn-reachable.
2. No record said which members were red for a known reason. A red sweep lists failed shards; which member failed, and whether someone owns it, has to be dug out of the logs.

Changes

The pull request that moves this index to mcpp 2026.9.28.2 carries both fixes:

• the alert job names the repository (GH_REPO) and also runs when mirror-cn-reachable fails;
• tests/known-red.tsv lists the members known to fail, per leg, with the issue that owns each; the job red members name their issues (tests/red_members.py) lists every failing member with its issue in the run summary, and fails when a failing member has none, or when the members' job failed and no member row can be attributed.

The pull request's full sweep is the first reading under the new rule. pangocairo passed on 2026-09-27, so the list starts empty; if it fails again it gets its own issue and a line.

[speak-agent]

Both defects are fixed in #483 (e9b80c5): the sweep alert sets GH_REPO and also runs when mirror-cn-reachable fails, and the new job red members name their issues lists every failing member with the issue that owns it (tests/known-red.tsv), failing when one has none. The full sweep with mcpp 2026.9.28.2 (run 36376460690) passed all 22 shards, and the new job read "No member failed." pangocairo passed there as on 2026-09-27, so tests/known-red.tsv starts empty. A future red sweep either names its issue or turns the summary red.