session: control the clocks in the slowest regression tests

[AbirAbbas]

What

Twelve internal/session tests spent about 123 seconds (sum of local medians, ten runs each) waiting on real wall clocks: watch tickers, the team traffic settle, the steer grace, a blocked phase listener, and a stalled checker's call share. They are a large share of the session leg in every pull request's touched packages run, and a real-clock window is the same shape behind the suite's load-only failures (#1664, #1339).

Each test now drives the same real commands, queues, landings and receipts, and waits on an acknowledgement or an explicit clock advance instead of on time passing. Sum of medians after: about 4 seconds.

Test	Before	After
TestASlowListenerNeverHoldsTheTurnThatIsTellingIt	42.5s	0.5s
TestTeamWakeTheLoopBreakerCountsTenMemberReplies	20.9s	0.2s
TestWatchQuietFiresWhenTheOutputStopsMoving	12.7s	0.2s
TestWatchMatchDeliversOnlyMatchingLines	10.6s	0.2s
TestWatchTicksNeverStartATurn	8.7s	0.6s
TestAStalledCheckIsAbandonedAndTheSecondCallAnswers	7.3s	0.4s
TestASteerLandsWhenAYoungBashCrossesTheGrace	7.0s	0.4s
TestWatchStopsAfterThreeIdenticalFailures	6.9s	0.4s
TestWatchChangeIsBaselineSilenceThenDeltaOnly	4.6s	0.4s
three task-baseline tests	2.3s	0.9s

Product code

Four private seams, each nil/false in production and falling through to the exact call it replaces: jobRegistry.watchTickWait, Agent.steerAfter, Config.teamWatchManual, and Config.auditTimeout → callWindow.timeout. No default, window, limit or person-facing behaviour changes, so no manual page changes.

Validation

• Before changing each test, the behaviour it pins was written down in one sentence.
• Mutation proof per test: the smallest product edit that breaks that behaviour was applied through a Go overlay, and the changed test fails on it; the product file was then verified byte-identical.
• -count=10 focused runs for every changed test; go vet, gofmt, make test-laws.
• make pr-ready on this branch: light gate green; the full session suite failed three tests that this branch does not touch (TestARefutedNodesRepairRoundRunsOnTheCarefulModel and TestTickingAnItemMovesTheRowToTheNextItem: "task 1 never finished" on a fixed 30 s wait; TestSpendRowsFollowThePlanNode: 3 spend rows for 2 calls) while the box was at load ~40. All three pass 3/3 on this branch and 3/3 on plain dev once the load dropped, and no helper they use is changed here. They are pre-existing load-sensitive tests.

Not changed

The two plandb CLI tests that run a real loop through bash (about 35 s each) are left alone: they are integration tests, and a faster version would stop running the real CLI.

🤖 Generated with Claude Code