Odin read and decoded its whole 3.24 MB store file about 30 times a second:
every refresh pass, every admission, every catalog query and every CAS retry
read it again. The working set now lives in memory under the write lease.
Deleted: CultCacheOdinTopologyStore (every file read and CAS against the
store, and its four CAS_ATTEMPTS retry loops, including the one in
reconcile), persist_generic_document's file CAS loop, the store half of
incarnation_keys, stored_snapshot's file read, and
prior_self_publisher_sequence's file read in RuntimeState::open. With one
writer under the lease an in-memory exchange cannot lose, so nothing retries.
Added: MemoryOdinTopologyStore, keyed (type, key), implementing the
OdinTopologyStore trait (still the mock point). It is loaded once, with the
locked pull_all, in try_activate after the lease is verified; legacy
retirement and the self publisher sequence are operations on the loaded set.
The trait's compare_exchange collapses into commit. Refresh reads Idunn's
projection once per pass (IdunnProjectionSnapshot) instead of once per
incarnation; admission reads it once per admission.
Not in this cut: the write policy. Each mutating operation still writes the
whole file immediately, now from memory (flush after an admission, a put or a
refresh pass that changed something).
Tests: the two tests that corrupted the store while Odin served now put the
bad record in the file at activation. New: rewriting the file behind Odin
changes nothing it serves; refresh passes and catalog reads succeed with the
file replaced by bytes that do not decode; a loaded store never reads its
file again.
Map: docs/write-pattern-cut.md, Cut 1.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>