Parent: #531 Related: #532, #478, #461
Problem
After #532 removed the accessor-reader serialization hotspot, the frozen shared transition-publication row scales through four lanes but still loses aggregate throughput at eight lanes. Current shared attribution shows the partitioned 8-lane invocation publishes 8,008 independent transition misses with zero Shape writer yields, two Object property-lock contentions, and no backing or element contention. A post-change collapsed sample instead places Shape.lookupState first among engine leaves.
The ordinary named-data mutation paths resolve the same immutable Shape/name state more than once while one Object property snapshot is already held:
- OrdinarySet checks the receiver state, then setOwnUnlocked repeats it;
- named Delete checks the current slot, then deleteTransition repeats it.
At 8 lanes this adds 16,000 redundant shallow probes to 8,000 exact add/read/delete operations.
Scope
- Carry a proven Shape.LookupState through the locked OrdinarySet and named Delete mutation helpers.
- Preserve one public lookup wrapper for callers that do not already hold a proven state.
- Add opt-in lookup request/comparison attribution if needed to prove the eliminated work without charging production-disabled paths.
- Keep Shape transition keys collision-independent and exact; do not change the skip-list publication protocol.
No-workaround rules
- No benchmark recognition, prepublished keys, skipped descriptor/prototype checks, weakened Proxy or accessor semantics, GIL, larger GC threshold, hidden lifecycle exclusion, or reduced work at higher lane counts.
- Do not modify stacks, queues, jobs, Promise/microtask behavior, or Thread/Worker lifecycle.
- Do not trade OOM atomicity, delete/re-add ordering, stable slots, same-name convergence, or moving-GC safety for fewer probes.
Acceptance
- Exact attribution proves one lookup per logically required state observation and identifies every eliminated duplicate.
- Base, anti-specialization variant, and same-parent contended rows retain frozen checksums and equal timed work.
- Exact-parent 1/2/4/8 evidence reports wall, CPU/instructions, dispersion, allocation, GC, RSS, and remaining synchronization.
- OrdinarySet inherited setters/non-writable descriptors, accessors, Proxies, arrays, module namespaces, typed arrays, delete/re-add order, and OOM rollback remain exact.
- Shape, Object, Array-prototype, and Proxy focused tests plus full units, no-GIL TSan, threadfuzz, moving/parallel GC, and exact-parent test262 pass.
- Any performance claim waits for clean checked-in raw evidence and report.