Fill memory reads after lifecycle filtering

[jordan-hivemind]

What changed

• Replace fixed candidate windows in structured fact and narrative memory reads.
• Add status-aware database indexes and a status filter field to full-text fact search.
• Apply lifecycle and business-time validity filters inside the Convex query before take(limit).
• Remove a helper that attempted multiple .paginate() calls inside one Convex function execution, which Convex rejects at runtime.
• Cover recent, core, historical, typed, and full-text reads with regression corpora that exhaust the previous candidate windows.

Why

The previous implementations fetched fixed candidate windows and filtered lifecycle state afterward. An account with enough retracted, superseded, scheduled, or expired records could therefore receive fewer results than requested even when additional retrievable records existed.

Convex also permits only one paginated query per function execution, so looping over .paginate() inside the existing query functions is not a valid repair. Query-level filtering makes take(limit) count retrievable rows while status-aware indexes keep the common current-only path efficient.

Validation

• pnpm test:once — 117 tests passed
• pnpm check-types
• pnpm lint — zero errors; ten pre-existing warnings
• pnpm build
• pnpm eval:recall — R@5 1.0, R@10 1.0, no blocking failures
• convex dev --once — schema/index backfill and function deployment succeeded

[vercel]

@jordan-hivemind is attempting to deploy a commit to the Pro Team on Vercel.

A member of the Team first needs to authorize it.