Three app crashes on 2026-07-24 shared one stack: SIGTRAP malloc abort
under BFCArena::Extend while embedding note chunks on the main thread.
The BFC arena doubles on every extension and never shrinks (1.6GB RSS
observed for the 25MB bge-small model), and unbounded batches — a 26KB
note embeds 21 chunks in one call — pushed its next multi-GB aligned
allocation into a fatal abort inside the Electron main process.

Disable the arena (plain malloc/free per tensor, memory returns to the
OS) and embed in batches of 4 to bound peak activation memory. Verified
in an isolated repro: 1.6GB → ~200MB stable, 10 full-note reindexes in
6.3s.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

sessions.json was only written on graceful quit, so a hard crash lost
every session with nothing to restore. Journal the same resumable
snapshot during runtime — every 15s plus immediately when a session
gets its Claude id — so a relaunch after a crash offers the normal
restore prompt, losing at most ~15s of session-list changes.

The journal skips content-identical writes and never overwrites a
previous run's saved sessions with an empty list before this run has
resumable sessions of its own, keeping the pending restore prompt safe
right after a crash relaunch.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>