execution, db: simplify cache publication and hot-path plumbing

[yperbasis]

Goal

Track independent simplifications and performance opportunities left after #23095. None of these is required for the correctness of #23095, and each can be reviewed separately.

Candidates

Treat missing commitment files as files end zero

Aggregator.beginCacheFilesPublication currently skips BranchCache publication when no commitment file is visible, while closeDirtyFilesNoReopen resets BranchCache explicitly.

Always call BranchCache.BeginFilesPublication, using end zero when the commitment view is absent. This would give file extension, lowering, and complete unmapping one publication protocol and remove the manual reset from closeDirtyFilesNoReopen.

Add a regression test proving that publication of an empty commitment-file view revokes the old view and clears incompatible entries.

Avoid copying every adaptive-pin miss snapshot

AdaptivePinController.PlanBlock copies the map returned by snapshotMisses, then deletes active contracts from one copy before selecting promotion candidates.

Keep the original snapshot as plan.observedMisses for abort restoration, and make candidate selection skip contracts that remain active. This removes one map allocation and one full copy per block. Tests must cover successful publication and abort restoration.

Derive cache views once per TemporalPutDel

temporalGetter derives its cache views once, but temporalPutDel calls the direct SharedDomains write methods. Writes with prevVal == nil can therefore derive the same transaction identity repeatedly while resolving previous values.

Store one cache-view pair in temporalPutDel and reuse it for previous-value reads. Measure the writer path before and after because this changes hot-path plumbing even though the transaction's state version and pinned files are stable.

Centralize typed CodeCache access

StateCache repeatedly asserts that its code-domain slot is a *CodeCache, although the constructor establishes that invariant. A private typed accessor can put the defensive check in one place and make the code-specific paths consistent. This is a small readability cleanup; it should not add another pointer to StateCache unless measurement justifies it.

Evaluate one lock for GenerationGate

GenerationGate uses publicationMu to order transitions and admissionMu to drain or block fills. One sync.RWMutex could represent both rules: fills take the read lock, while initialization, reset, close, canonical publication, and backing-file publication take the write lock.

This would reduce the lock-state proof and remove unlock/relock choreography. Do not adopt it without contention benchmarks: Go writer preference may make fills block earlier when publications queue, even though cache hits remain lock-free.

Non-goals

• Do not merge StateCache and BranchCache generation gates. Their backing-file identities change independently, so a shared gate would cause unnecessary invalidation.
• Do not add another generic publisher layer. GenerationGate and CanonicalPublisher already share the common lifecycle; state updates and branch/adaptive-pin updates remain meaningfully different.
• Do not remove explicit unwind, file-publication, and concurrency scenarios merely to reduce test diff size.

Suggested order

1. Normalize the empty commitment-files publication.
2. Remove the adaptive miss-map copy.
3. Centralize CodeCache access.
4. Benchmark cached TemporalPutDel views.
5. Benchmark the single-lock GenerationGate design before deciding whether its simpler state machine is worth the scheduling change.

[VBulikov]

yperbasis added Imp2 as Importance