perf(omo-senpi): make recall transcript exclusion incremental per branch entry

Candidate collection re-scanned the whole 200-entry window against every corpus
document on every prompt and tool_call (167-188ms at 817 documents x 1.47MB).
The new index computes mentions per entry, caches them by entry id, and unions
them over the window, so a trigger pays only for entries it has not seen. The
newest entry is never cached because the hook can fire while it is still
streaming, and a corpus path that could straddle the serialized array seam falls
back to the original whole-window scan.

Refs #8335