[Soushi888]

SoushAI analysis. Drafted by Soushi's AI assistant, reviewed and posted by @Soushi888.

Intent

Two fixes for #142, one per problem in the issue. The nondominium Sweettest target ran close to 30 minutes, so it now runs as two halves, and the job gets an explicit 45-minute timeout. The e2e "empty lobby" precondition kept failing because Playwright retried the serial core-flows story against the same conductors; that story no longer retries, so the first real failure is the one reported. Measured on this PR's first CI run, the resource split did not pay off and was dropped: resource stays one target.

Changes

Sweettest shards (.github/workflows/build.yml). The matrix becomes include entries, and nondominium runs as two halves (a, b) of the same [[test]] target, selected by libtest name filters in the run step:

• half a passes the filter list (substring match, any filter selects);
• half b passes --skip for exactly the same list.

Every test runs in exactly one half, and a new test can never be dropped: it lands in b unless its name matches a filter. Rebalancing means moving one filter. No test file changes.

target	half a	half b
nondominium	lifecycle, transition history, hibernation, deprecation, invalid transitions (9)	create, discovery, property regime, membership (9)

timeout-minutes: 45 on the sweettest job: without it the runner default is 360 minutes, so a hung conductor held a runner for six hours. A hang now names itself in the job annotation.

Job names change from sweettest (nondominium) to sweettest (nondominium, a) and (nondominium, b). dev has no branch protection or rulesets (checked through the API), so no required check is orphaned.

Empty-lobby precondition (ui/tests/e2e/specs/core-flows.spec.ts, ui/tests/README.md). test.describe.configure({ retries: 0 }) inside the core-flows serial story, with a comment naming the cause, and a line in the e2e README.

Decisions

• The resource split was measured and dropped. On this PR's first run (37550976485) the resource halves took 5.5 and 20.1 min, against 12.7 to 15.9 min for the whole target on dev (runs 36675002358, 39905805052, 37839257499). Half b alone outran the whole target and became the path that gates e2e, for the cost of an extra runner. The nondominium halves came in at 17.0 and 14.8 min against 27.5 to 29.2 min whole, a real gain, so that split stays.
• The 59m51s resource failure was not a timeout. Its annotation on run 32910203007 reads "The hosted runner lost communication with the server", runner starvation. The timeout does nothing for that; it stays the only instance on record.
• The leaked lobby group is the suite's own retry. In all three failing runs (32914253920, 32916185361, 32917489986) attempt 1 passed the empty-lobby test and failed later, at lifecycle history panel present. Playwright retries a describe.serial group as a whole against the same conductors, so each retry started with attempt 1's E2E Circle group still in the sidebar, and the precondition error replaced the real one. test(v0.1.0): harden e2e and Sweettest coverage for the peer-mesh release #127's clone-cell teardown is not involved.
• Not taken: tearing down agent 1's clone cells at the start of a retry. The lobby DNA would still hold the previous attempt's profile and NDO announcements, so a retry would still not start clean, and would fail somewhere less legible. What we lose is a retry for a transient failure in the story's first tests; those are still reported accurately.
• multi-agent.spec.ts is also describe.serial with the same retry exposure. Left alone because the issue does not name it; a candidate for the same one-line change.

How to test

• CI on this PR: the checks list sweettest (nondominium, a) and (nondominium, b), and one sweettest (resource). Compare their durations with gh run view <id> --json jobs.
• actionlint .github/workflows/build.yml: exit 0.
• Split coverage, real libtest semantics: cargo test --package nondominium_sweettest --test nondominium -- --list with each half's filters inside nix develop gives 9 + 9 = 18 tests, 0 overlap, union equal to the full target.
• Retry masking, red then green, in a throwaway Playwright 1.61.1 project with retries: 2 and a file standing in for conductor state: without the change the real failure is reported as flaky and the precondition error fails on both retries, the exact CI pattern; with retries: 0 the real failure is the only one reported.
• playwright test --list from ui/ with CI=1: 23 tests in 2 files.

Related

• Related ci: the resource Sweettest shard is nearing the job timeout, and the empty-lobby precondition flakes on leaked state #142 (not Closes: the issue asks for the e2e side to be watched over several runs before it is closed)

Pre-open review

Run 2026-10-27 through PRReview at db3d80c, re-run after the fixes at dad7fb0 (rebased onto dev after #157).

• Description: rewritten to the repository template (Intent, then Changes, Decisions and How to test folded, Related), with Related #142 under Related.
• Merge state: rebased onto dev (808fcd7), no conflicts.
• Code findings: the resource split cost more than it saved on the PR's own measurement; dropped in dad7fb0. The filter design was checked against the test names: 9/9, no accidental matches.
• Documentation: passes; ui/tests/README.md documents the retry opt-out, and the workflow comments document the split with its measurements. No file under documentation/ describes the CI matrix.
• Tests: CI and test configuration only; the PR's own CI run is the measurement.
• Left open: the same retries: 0 for multi-agent.spec.ts.