Record failed detector carry-forward gate

[Labeeb2339]

What changed

• add an opt-in persistent detector with three-token persistence and 64-token cooldown
• add a fixed 64-token event-gated retention intervention while preserving the historical default policy
• run the preregistered 30-seed synthetic sanity sweep and check in all per-seed evidence
• add deterministic replay coverage and document the failed carry-forward decision

Result

The detector found 30/30 known midpoint shifts with a six-token median delay and zero stationary false triggers. However, the intervention increased first-64-token post-shift modeled link bytes by a median 11.52%; the paired-bootstrap 95% interval was [+10.78%, +12.64%]. The preregistered traffic gate failed and carryForward is false.

No threshold, hold interval, or weight was retuned after observing the sweep. This PR stops the current JSD-gated retention design rather than presenting the timing result alone as a win.

Validation

• npm run check
• 27 unit tests and 2 rendered-app tests passed
• lint, typecheck, production build, and production dependency audit passed locally
• a second full sweep produced the identical evidence SHA-256
• every checked replay preserved router selections, issued no prefetch, and reported zero semantic routing changes