fix(db): bound cohere's writer-lock window — split the maintenance transaction and chunk the decay pass

cohere previously ran promotion, a full-table decay UPDATE, link building,
and archive inside ONE BEGIN IMMEDIATE, so the writer-lock hold grew
linearly with store size (measured 0.34s @100k / 0.68s @200k release on a
fast dev box; 4.45s @100k on the #400 deep-dive box) and crossed the
default 5000ms busy_timeout just past ~130k entities — every concurrent
remember failed SQLITE_BUSY during maintenance runs.

The decay pass needs no atomicity with promotion or link building, so
cohere now runs three bounded lock windows: promotion; the ×0.95 decay
groom chunked at COHERE_DECAY_CHUNK_ROWS (1000) rowids per drop-safe
IMMEDIATE transaction with a 2ms inter-chunk yield (SQLite's coarse busy
handler otherwise starves waiters across back-to-back chunks); and
link+archive (candidate_budget-bounded, links read-modify-write kept under
one IMMEDIATE tx so the #388 clobber analysis still holds). The new
non-atomicity boundaries are documented at the split site; every step is
idempotent and each transaction stays drop-safe.

Longest single hold @200k: 0.68s -> 0.09s; write work per hold is now
bounded by the chunk size, and the residual linear component (the
promote/archive full-table read scans) is ~7.5x shallower. Preserves the
carry the same epsilon guard) and cohere's documented standalone ×0.95
decay semantics (deliberately NOT deduplicated into decay_tick's
Ebbinghaus recompute; only the chunking structure is aligned).

Regression tests: chunk-transaction count for a store above the chunk
size; no-op skip write-count for floored rows; a concurrent writer with a
~250ms fine-grained wait budget succeeding during cohere @150k (fails 3/3
on pre-fix main, passes 3/3 post-fix); plus an #[ignore] lock-window
measurement harness (writer-lock hold probe over a seeded store,
MIMIR_COHERE_MEASURE_ROWS). Also adds a store-size ops note for
MIMIR_BUSY_TIMEOUT_MS at the pool-config site.

Closes #400

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

Author: tcconnally

CHANGELOG.md

@@ -48,6 +48,23 @@ All notable changes to Perseus Vault (formerly Mimir/Mneme) are documented here.
   truncation note when capped. Edges dangling outside the returned node set
   are dropped (previously the unscoped path emitted edges to archived/deleted
   targets that the renderer couldn't resolve).
+- `mimir_cohere` no longer holds one writer lock for the whole grooming pass
+  (#400). The single BEGIN IMMEDIATE previously spanned promotion, a
+  full-table decay UPDATE, link building, and archive — a lock window linear
+  in store size (~4.4s @100k entities) that crossed the default 5000ms
+  `busy_timeout` just past ~130k entities, so concurrent `remember`s failed
+  SQLITE_BUSY during every maintenance run. cohere now runs three bounded
+  windows: promotion, a decay pass chunked at 1000 rows per drop-safe
+  transaction (with a 2ms inter-chunk yield so waiting writers can actually
+  acquire the lock), and link+archive. Longest single hold measured @200k
+  entities (release, ~450B rows): 0.68s → 0.09s; the write work under any
+  one lock is now bounded by the chunk size, and the remaining linear
+  component (the promote/archive full-table read scans) is ~7.5x shallower
+  than before. Preserves the #399/#405 no-op write skip (floored
+  rows are not rewritten), cohere's documented ×0.95 standalone decay
+  semantics, and per-transaction drop-safety (#388); the run stays correct
+  under interleaved writers, and the new non-atomicity boundaries are
+  documented at the split site.
 
 ## [2.14.0] - 2026-07-02