Erase async task body types at RuntimeHandle::spawn so SQLite owner loops
share Tokio scheduler code. Keep blocking SQL work and results concrete,
and retain the existing task admission, failure collection and shutdown.
This boxes each submitted async task once; normal SQL commands do not gain
an additional box. The original Linux x86_64 release profile, followed by
strip --strip-all, decreases from 111026760 to 110965320 bytes (60 KiB).
Validation: cargo fmt --all; full workspace Clippy with all targets/features
and -D warnings; Nextest 22154 passed, 20 skipped; CDP smoke 92/92 on both
Moli and actual Chrome 145.
Discard failed Cookie batches without requeueing, following Chromium's commit
behavior. Preserve the live canonical jar and the normal 30-second/512-operation
schedule for new mutations, and report persistence errors without replaying the
old batch on flush/close.
Remove backoff state, error classification, and per-call failure admission.
Retain explicit-flush recovery only for localStorage, whose absolute area
metadata must stay paired with its failed entry deltas.
Cover idle failures, subsequent batch deadlines and operation thresholds,
non-replaying flush/close, Cookie jar versus SQL state, and actual SQLITE_FULL
recovery by explicit localStorage flush.
Keep the latest Put or Delete for each key rather than eliding transient-row
deletes. Remove existed from Cookie and localStorage deltas and Cookie mutation
observers, then remove the shared uncertain-commit correction hook. Preserve
Cookie Touch composition and atomic, idempotent retry behavior.
Replace optimization-specific assertions with actual SQL replay tests covering
a published put followed by deletion and clear followed by a newer put.
Preserve SQLite error codes and retry atomic batches after transient resource,
locking, and I/O failures. Back off from one second to a capped sixty-second
interval without letting new mutations bypass the delay. Successful commits
restore localStorage admission; permanent errors still require explicit retry.
Verify bounded scheduling with virtual time and real SQLITE_FULL recovery
without a second flush, including restart persistence and new writes.
Replace fixed worker shards and Lane locks with one async worker and one FIFO owner per database. Move synchronous SQL, connection work and catalog operations to the blocking pool. Drain accepted work and await final owner receipts before retirement or runtime teardown.
Use a single maintenance task with a short idle grace and at most one pending check per database, so slow owners cannot stall later rounds. Retry transient open failures in the same prepared generation while preserving terminal missing-file checks. Cover lifecycle races, low-FD creation and recovery, and single-worker host shutdown without new shared resource budgets.
Reserve bounded cleanup metadata and one control slot before fencing an owner. Close each database on its SQL lane, retire generation directories, cancel accepted factory requests and only then allow a new generation. Return logical connection retirement capacity at the native close receipt even when callers retain closed wrappers.
ldm0