[freeqaz]

Problem

The static-egress test "sustained sequential operations outlive replay capacity without waiting for JWT expiry" runs 4200 sequential operations. The literal was sized to pass the original 4096-entry replay table. Since #3281 the table holds 8192 entries and one Sandbox has a 2048-entry share, so a single Sandbox reaches its share long before the table. The test now guards the per-Sandbox replay share, not the table. It is also the slowest case in Checks and Conformance 1: usually 43-70 s in CI, and 93 s and 124 s on slow runners (180 s timeout).

Change

• Derive the loop count from the exported constants: REPLAY_SANDBOX_SHARE + 64 operations (2112), with the same 25 ms clock step.
• Assert the test's premises, so a later constant change cannot quietly make it vacuous:
  • one 30 s horizon of entries at that step (1200) fits under the share;
  • the share is below the table;
  • OPERATION_CAPACITY is below the loop count.
• Add a comment naming which bounds the test exercises. The installation-wide table's expiry stays covered by "one Sandbox's replay share bounds it without refusing other Sandboxes".

Test-only change; the test name is unchanged, so the suite manifest is untouched.

Verification

Mutants of packages/occ/src/credential-proxy.ts, run against the shortened test (Node 24, 2 CPUs):

Mutant	Result
pruneReplay returns at once (no replay expiry)	fails at operation 2048
table entry deleted, Sandbox replay share never returned	fails at operation 2048
releaseCharge never returns the Sandbox operation share	fails at operation 256
releaseCharge never returns activeOperations	fails at operation 1024
share returned, table entry never deleted	passes this test before and after the change (4200 < 8192); fails "one Sandbox's replay share bounds it without refusing other Sandboxes"

Timing:

• Local, 2 CPUs (taskset), before and after run concurrently, 4 rounds each: before 38.8-48.7 s (median 42 s), after 19.1-25.5 s (median 21 s).
• CI, Checks and Conformance 1 on this head: the case took 37.9 s and its file 54.5 s. The file no longer ends the lane: the last file ends at 66.8 s.
• Latest main (9db4478): the case took 53.6 s, and the file ran 65.2 s and ended the lane at 68.1 s.
• The slow main run on f2a5b4b: the case took 92.7 s, and the file ran 106.5 s and ended the lane at 110.1 s.

Pin: the whole file 20/20 at 2 CPUs (5 concurrent loops on separate CPU pairs; 78 tests per run, file 30-44 s, this case 19.9-29.3 s).

CI Required passed on 5ff8f1e; First Agent Smoke was skipped by the test-only path filter. An internal review approved with one comment nit, now applied.

Risks

Low; test-only. The shortened loop still runs 64 operations past the share, each of which needs an entry released by its horizon. Regressions that only show after many more horizons would not be caught, but no current bound depends on more than one share's worth of operations.

[clawsweeper]

🦞👀
ClawSweeper picked this up.

Pull request received. I will update this pull request when review starts.