test(fallback): fix flaky cooldown-window timing assertion

[teknium1]

Summary

test_non_retryable_exhaustion_arms_cooldown is a freshly-merged flaky test (added in #53909) that fails reliably on loaded CI workers. It captured before = time.monotonic() ahead of three _try_activate_fallback() calls, then asserted the armed cooldown was <= before + 5.0 + 1.0. The cooldown is set relative to the final call, and the activation work (agent init, resolve_provider_client) takes >1s under load — so the before-anchored bound overshoots and fails.

Anchoring the upper bound to after (captured once the cooldown is armed) makes the assertion immune to activation latency without weakening what it verifies.

Changes

• tests/run_agent/test_24996_fallback_exhaustion_cooldown.py: capture after = time.monotonic() after the activation block; assert cooldown <= after + _FALLBACK_EXHAUSTED_COOLDOWN_S + 1.0.

Root cause

cooldown == final_call_time + 5.0, and after >= final_call_time, so cooldown <= after + 5.0 + 1.0 holds regardless of how slow the three activation calls are. The old before-anchored bound did not — any activation latency >1s leaked straight into the slack.

Validation

	Before	After
CI slice 8/8	assert 230.21 <= 229.73 FAIL (observed twice; also 905.80 <= 905.31)	bound tracks after, no drift
Local pytest (file)	—	4 passed
Local stress (target test ×5)	—	5/5 pass

Still proves the short 5s window vs the 60s rate-limit one — far below the +50s discriminator the sibling test_rate_limit_exhaustion_keeps_60s_cooldown relies on. Test-only, +7/-1, no logic change.

Infographic

[github-actions]

🔎 Lint report: fix/flaky-fallback-cooldown-timing vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 11602 on HEAD, 11604 on base (✅ -2)

🆕 New issues (1):

Rule	Count
invalid-assignment	1

First entries

tests/run_agent/test_credits_notices_toggle.py:76: [invalid-assignment] invalid-assignment: Object of type `None` is not assignable to attribute `_credits_session_start_micros` of type `int`

✅ Fixed issues (2):

Rule	Count
unresolved-attribute	2

First entries

tests/run_agent/test_credits_notices_toggle.py:76: [unresolved-attribute] unresolved-attribute: Unresolved attribute `_credits_session_start_micros` on type `AIAgent`
run_agent.py:3014: [unresolved-attribute] unresolved-attribute: Object of type `Self@get_credits_spent_micros` has no attribute `_credits_session_start_micros`

Unchanged: 6087 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.

[tonydwb]

Code Review Summary

Verdict: Approved

Clean test fix. Anchoring the upper bound to after (captured once the cooldown has actually been armed) rather than before eliminates the wall-clock jitter that caused CI flakes. The rationale in the comment is clear. Note: this PR also includes an infographic PNG, which is fine.

---

Reviewed by Hermes Agent

[teknium1]

Superseded by 88b3d86 (test: de-flake SIGKILL-tree, compression-tip resume, and fallback-cooldown tests), which landed on main and fixes the same flake in the same test via a wider +30s slack. Same root cause we diagnosed here (tight +1.0s upper slack false-failing on loaded CI runners). Closing as redundant — the test is no longer flaky on main.