test(decay): deflake reinforced-memory guard against coarse clocks

[Coding-Dev-Tools]

Problem

tests/test_decay_idempotent.py::test_reinforced_memory_is_not_decayed_that_interval failed intermittently in full-suite runs on this Windows host:

AssertionError: assert 2.1281674200679393e-09 < 1e-09

It always passed in isolation.

Root cause

The test simulated reinforcement with last_access = now_ts() immediately after a decay pass had written last_decay = now_ts() as its anchor. On hosts with coarse clock resolution (consecutive time.time() calls can return the same value — observed on Windows), last_access == last_decay, so the production guard

if last_access > anchor:   # strict >

does not fire, and the pass legitimately decays one clock tick's interval. At a 7-day half-life, one ~1 ms tick decays stability by ~2.1e-9 — just over the test's 1e-9 bit-identity threshold. This is a test race against wall-clock granularity, not a recurrence of the compounding-decay regression the test exists to pin.

Fix (test-only)

• Set last_access = now + 1.0 (one second past the pass anchor) so the guard condition holds deterministically on every host, regardless of clock resolution.
• Extend coverage: assert the guard across 50 rapid subsequent passes, not just the first one — that is the actual invariant (a reinforced memory is never decayed again until new real time passes).

No production code changed. The strict > guard is correct: with last_access == anchor on a first pass (last_decay IS NULL, anchor derived from last_access), the 10-day interval must decay; changing it to >= would break first-pass decay.

Verification

• tests/test_decay_idempotent.py passes 8/8 consecutive runs (previously flaked ~1 in 3 full-suite runs).
• 20/20 same-tick stress trials: stability bit-identical across rapid passes.
• Full offline gate: 4503 passed, 39 skipped, ruff clean, pyright 0 errors, all eval gates green.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.