maintain() calls rewriteTranscriptEntries() which rewrites the JSONL via
branch-and-reappend, changing file size, mtime, and entry IDs. But the
conversation_bootstrap_state was only updated during bootstrap(), leaving
a stale checkpoint after every maintenance rewrite.

On the next gateway restart, bootstrap saw the stale checkpoint:
- Fast path 1 (exact size+mtime match) failed (file shrank)
- Fast path 2 (append-only, size > stored) failed (file shrank)
- Fell through to reconcileSessionTail() full content-based reconcile
- On conversations with many identical (role, content) pairs, the
  occurrence-counting anchor matched at the wrong position by coincidence
- Result: thousands of duplicate messages imported in one second

Three fixes:

1. Checkpoint update after maintain(): After a successful JSONL
   rewrite, update the bootstrap checkpoint so the next bootstrap()
   hits the fast path instead of falling through to reconcile.

2. messageOnly skip in readLastJsonlEntryBeforeOffset(): The function
   now accepts a messageOnly flag so it skips non-message JSONL entries
   (cache-ttl, tool-result, session-meta) when computing the tail
   entry hash for checkpoint comparison.

3. Import cap in reconcileSessionTail(): If reconcile would import
   more than 20% of the existing DB message count (minimum 50), abort
   and log an error instead of blindly importing.

Root cause of 19+ bootstrap flood events across March 29 - April 5, 2026.

Fixes Martian-Engineering#271
Relates to Martian-Engineering#268
Relates to Martian-Engineering#276

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Keep maintain() checkpoint writes aligned with bootstrap state semantics by normalizing the stored mtime, and prevent capped reconcile aborts from advancing bootstrap state as though the transcript had been fully processed. Add regression coverage for the unchanged maintain->bootstrap fast path and the capped reconcile retry path, plus the required patch changeset.

Regeneration-Prompt: |
  Address the two review findings on PR 280 without broadening scope. Make bootstrap checkpoint updates after transcript maintenance use the same timestamp semantics as the existing bootstrap path, and ensure the replay-safety import cap in reconcileSessionTail does not mark the transcript as fully processed when it aborts. Add targeted regression tests proving maintain() leaves the next unchanged bootstrap on the fast path and that a capped reconcile preserves the stale checkpoint so a later retry is still possible. Include the required patch changeset and rerun the repository's relevant test and packaging gates before pushing back to the contributor fork.