fix: six confirmed audit findings (token lifecycle, msg-drop, recall scaling)

[saucam]

What

Fixes the six confirmed, high-severity findings from a full-codebase audit. Each was adversarially verified against the actual source before fixing (one skeptical verifier per finding, quoting the code). The scarier "stuck-status / race" claims from the first pass were refuted on verification and are deliberately not touched.

Fixes

#	Area	Problem → Fix
1	session.ts	Dropped message after interrupt-then-fast-send. The consumer finally nulled #inputQueue/#consumerTask without the identity guard it already used for #query/#abortController; an un-awaited interrupt() + fast send() let the stale loop clobber the new loop and the next push was silently dropped. → Loop-local queue/task snapshot + identity-guarded nulls.
2	auth.ts · server.ts · types.ts	Token never re-verified after handshake → an open socket honored an expired/revoked token forever. → Carry exp into AuthContext, reject missing/expired exp in verifyToken (60s skew), close 4003 on a per-message expiry check.
3	web/lib/ws.ts · connection.ts	Web reconnect replayed the dead JWT forever. → getToken() supplier re-exchanges the stored zid_sk_ key on every (re)connect; falls back to the last token. +2 tests.
4	memory/store.ts	Vector recall had no cache despite the comment — every recall() re-read + re-decoded all embeddings and brute-forced cosine. → Memoize the decoded matrix per workspace; invalidate on insert-with-embedding / setEmbedding.
5	memory/engine.ts	Memory init was all-or-nothing — an embedder download hiccup nulled the whole engine (also killing FTS recall + usage persistence). → try/catch embedder.init(), run FTS-only on failure; recall + embed-pump guard on a ready flag.
6	session-manager.ts	Unbounded session resume blocked startup (#6). → Newest-first sort, cap to 50, time-box to 20s, log what was left on disk.

The #2 + #3 pair closes the token-lifecycle gap that made the documented "kill a token → immediately lose access" guarantee false on long-lived connections. (Instant revocation of a still-valid token still needs a periodic re-verify — tracked as a follow-up.)

Verification

• bun run typecheck clean.
• 554 daemon tests pass; 95 web tests pass (vitest). (bun test globbing the web vitest files reports false fails — those are the known vi.advanceTimersByTimeAsync-under-bun incompatibility, green under cd web && bun run test.)
• bun run build produces cli.js.
• Live smoke against the dev daemon: restarted on this code, the new per-message exp check accepted a valid dev token, auth.ok + session.list succeeded, clean boot.

Closes #6.

The remaining medium/low findings (input schema validation, WS payload/backpressure limits, synchronous=NORMAL, shutdown ordering, retention/pruning, aud default, empty-tenancy reject, search-snippet tenant scoping, dead retry.ts, Telegram client pruning) are filed as separate issues.

🤖 Generated with Claude Code