The snapshot store's bounded-retry test proves the deadline by attempts, not by 40 ms of wall-clock slack #3148

The bounded-retry test in test/snapshot-store.test.ts now checks attempts and budgets instead of wall-clock time. The file passed 20 runs in a row, and the new test took about 350 ms when run alone. The commit is pushed. I couldn't run bun x tsc --noEmit: it failed in this sandbox with "bun is unable to write files to tempdir: ReadOnlyFileSystem". The forge's check will run the type check.

• What changed: the test used to assert took < 300 + 40 on real time. It now uses a fake clock that only the injected sleep advances, so nothing depends on scheduler latency.
• What it checks:

- The store makes exactly 5 attempts, and the error ends "failed after 5 attempts: no answer in 150 ms". - The budget left at the start of each attempt is [300, 290, 270, 230, 150]. - Each attempt's share is floor((left − waits still to come) / (5 − n)), which works out to [30, 37, 50, 75, 150]. The test computes these from the recorded budgets rather than reading them from the code.

• Removed: the took < 340 and took > 150 assertions, along with the real-time reads they needed.
• Not touched: withRetries — sleep was already injectable, so no clock seam was needed.
• Left out: the per-attempt timers still run on real time (the last stalled attempt waits about 150 ms). The test doesn't measure that, so only the test's duration is affected, not its result.

To try it: for i in $(seq 20); do bun test test/snapshot-store.test.ts || break; done.

Closes #3143

The bounded-retry test in test/snapshot-store.test.ts now checks attempts and budgets instead of wall-clock time. The file passed 20 runs in a row, and the new test took about 350 ms when run alone. The commit is pushed. I couldn't run bun x tsc --noEmit: it failed in this sandbox with "bun is unable to write files to tempdir: ReadOnlyFileSystem". The forge's check will run the type check.

• What changed: the test used to assert took < 300 + 40 on real time. It now uses a fake clock that only the injected sleep advances, so nothing depends on scheduler latency.
• What it checks:

- The store makes exactly 5 attempts, and the error ends "failed after 5 attempts: no answer in 150 ms". - The budget left at the start of each attempt is [300, 290, 270, 230, 150]. - Each attempt's share is floor((left − waits still to come) / (5 − n)), which works out to [30, 37, 50, 75, 150]. The test computes these from the recorded budgets rather than reading them from the code.

• Removed: the took < 340 and took > 150 assertions, along with the real-time reads they needed.
• Not touched: withRetries — sleep was already injectable, so no clock seam was needed.
• Left out: the per-attempt timers still run on real time (the last stalled attempt waits about 150 ms). The test doesn't measure that, so only the test's duration is affected, not its result.

To try it: for i in $(seq 20); do bun test test/snapshot-store.test.ts || break; done.