Why one issue, not two

Both items are followups to v0.9.3's bootstrap externalization landing (#510) and its diagnostics companion (#512). One concerns a perf regression risk introduced by the new ingest path; the other is an observability lifetime bug in the diagnostic that exists to surface bootstrap problems. Filing together so the bootstrap-import area can be considered as a single follow-up surface.

Sub-item D1: Per-message ingestSingle loop in #510 may regress first-time import latency

• Repro hypothesis: Sessions with thousands of historical messages (typical for power-user transcripts that triggered Bootstrap/reconcile can re-import huge historical messages and create multi-million-token conversations #489 in the first place) now incur N×ingest write traffic in a single withTransaction, where v0.9.2 used bulk createMessagesBulk + appendContextMessages.
• Evidence: PR fix: apply ingest protections during bootstrap import #510 replaced the bulk path with a per-message ingestSingle loop. Inside one withTransaction call, summary-store and conversation-store writes happen per message rather than once. PR carries no benchmark or perf budget. The exact audience that hit Bootstrap/reconcile can re-import huge historical messages and create multi-million-token conversations #489 (multi-million-token transcripts) is the audience most likely to feel any regression here.
• Suggested fix: Retain bulk insert as a fast path when no message in the batch would trigger an interceptor (cheap pre-scan over content blocks for image/file/oversized markers); fall back to per-message ingestSingle only when the pre-scan detects something. Alternatively, chunked transactions (e.g. 100-message chunks) so a long transaction does not block other DB activity.

Sub-item F10: recentBootstrapImportsByConversation has no TTL

• Repro hypothesis: A conversation that imported 7 times during early debugging will keep showing the "7 imports" diagnostic tag long after those imports have ceased to be relevant — until the LRU eviction at 100 conversations kicks in.
• Evidence: PR fix: add overflow assembly diagnostics #512's recentBootstrapImportsByConversation map only evicts on LRU pressure (capacity 100). There is no per-entry TTL. For a long-lived gateway the stale tag trails the conversation indefinitely.
• Suggested fix: Add a per-entry timestamp; treat entries older than e.g. 1h as expired on read; OR clear the entry once a conversation has had a successful round of subsequent normal turns without re-import.

Suggested PR scope

One PR adding a pre-scan fast path to the bootstrap ingest in src/conversation/bootstrap.ts (or wherever #510 landed the new loop), restoring bulk insert when safe and keeping ingestSingle for the externalization-required path. Same PR adds a TTL field to the recentBootstrapImportsByConversation entry shape and prunes on read in the diagnostic accessor. Both changes are localized and testable without standing up a full bootstrap fixture: pre-scan logic gets a unit test against synthetic message arrays, TTL gets a unit test by injecting a clock.

Cross-links

• Closes Bootstrap/reconcile can re-import huge historical messages and create multi-million-token conversations #489 was the original re-import root-cause that fix: apply ingest protections during bootstrap import #510 addresses.
• fix: apply ingest protections during bootstrap import #510 — bootstrap externalization PR introducing the per-message loop.
• fix: add overflow assembly diagnostics #512 — diagnostics PR introducing recentBootstrapImportsByConversation.
• Need overflow diagnostics showing top LCM prompt contributors and duplicate clusters #495 — adjacent diagnostic pulse work this should integrate cleanly with.