[codex] Stabilize rolling-window test fixtures

[rushillshah]

Closes #23

Scope

• Derive rolling-window test fixtures from the test run instead of fixed calendar dates.
• Keep Gmail, relationship extraction, role inference, and corrupt-JSON coverage inside the configured 30-day window.
• Add an explicit 29-day/31-day boundary regression so production window behavior remains covered.

Root cause

Four integration fixtures used June 2026 timestamps, so they aged outside the 30-day import/extraction windows as the wall clock advanced.

Validation

• npm run check
• npm test (70 passed)
• Focused rolling-window regressions (5 passed)

Risk

Low. Test-only change; production defaults and code paths are unchanged.

[rushillshah]

Maintenance blocker (2026-07-13): CI passes on Node 20 and 22 and the branch is mergeable, but this PR is still draft and review-required. Next action: mark ready/request review, then squash-merge; afterward rerun PR #24 CI.