test: 86 timeouts still sit at 120s, and an 85.2s case is about one bad runner draw from red

[wormeyman]

Follow-up to #203, which fixed one file and deliberately left the rest.

What #203 established

previewAgreement.spec.ts's Vulcanus rock and cliff coverage test failed CI with Error: Test timed out in 120000ms at 150.5s, on a docs-only PR. Its own work had not changed. Across five consecutive runs:

tree	duration	result
main, before the scrap work	69.6s	pass
#202	90.1s	pass
main, after #202 merged	108.8s	pass
#203	150.5s	timed out
#203, at the new 300s budget	139.7s	pass

That file now uses }, 300000). Nothing else was touched.

Why this is not finished

86 annotations across 29 files still carry }, 120000), and the ceiling is per-test and hand-written, so a shard rebalance changes which tests sit near it.

The nearest to the edge is vulcanusCliffRejectionStage.spec.ts:

the post-filter model predicts 1,662 survivor-keeps-edge cases and the game shows 0 - 85.2s

Run-to-run spread on these runners measures about 40%, which was established directly rather than inferred: #202's run and #203's run share the same 210 spec files, so vitest hands them identical shards, and the jobs still came in at 259 / 351 / 389 / 133 against 366 / 273 / 327 / 137 - shard 1 up 41%, shard 3 down 16%, and the binding shard changing identity from 3 to 1.

85.2s x 1.4 is 119s. That is inside the ceiling by under a second.

Options, in the order they were weighed

1. Sweep all 86 to 300000. 29 files, mechanical, removes the class of failure. Costs a large diff, and slows a genuine hang to 5 minutes before it reports.
2. Raise only the files with a case above ~60s. Smaller diff, same protection where it matters, but needs the per-test scan re-run whenever the suite grows.
3. Delete the annotations and raise the global testTimeout from 30s. One definition instead of 91, which is how CLAUDE.md frames why the annotations exist at all. Largest behavioural change, and it weakens the budget for every fast test too.

Deliberately not done in #203, which was a docs PR. No action is needed until a shard goes red on a timeout; CLAUDE.md's testTimeout section already names this file as the first place to look when one does.

What is NOT the fix

Retries are banned here and the reason still holds - nothing in this suite is nondeterministic, so a retry would only hide a real regression. And splitting the heavy file is not it either: #203 measured that and recorded why in CLAUDE.md's CI section.

[wormeyman]

Re-counted 2026-09-07: the figures in this issue have expired, and the file it names is gone

On main at ab55bc6:

• }, 120000) is now 25 annotations across 9 files, not 86 across 29. git grep -c '}, 120000)' -- 'test/*.spec.ts':

  file	120s annotations
  test/renderWorkerEngine.spec.ts	7
  test/cellSurvey.spec.ts	5
  test/elevationRenderRequest.spec.ts	4
  test/viewNormalisation.spec.ts	3
  test/cliffOreDirection.spec.ts	2
  test/tier2Coverage.spec.ts	1
  test/wasmFulgoraRenderParity.spec.ts	1
  test/wasmIslandFinderParity.spec.ts	1
  test/wasmNauvisParity.spec.ts	1

• }, 300000) is 85 across 9 files.

• test/vulcanusCliffRejectionStage.spec.ts, the "one bad runner draw from red" file, no longer exists - Delete the ported TypeScript specs (#227) #360 deleted the investigation specs, and Port phase 7: cut over and delete the TypeScript math #227 / The nine surviving src/noise/ math modules are held by the expression layer, not by Fulgora's renderer #371 deleted the TypeScript noise math the heavy tests exercised.

So the nearest-to-the-edge test has not been re-derived since the suite changed shape. Per CLAUDE.md, that needs a per-test read off a CI run, not a grep. Leaving this open as the record that the question is unanswered, but nothing in the body above describes the current tree.