perf: reduce subscription churn and retained registration storage

[wallstop]

Part of #497 and the SubUnsub outcome in #414.

Inputs and dependencies

• perf: calibrate the IL2CPP measurement system and attribute dispatch cost #500 coordinates the measurement foundation.
• perf lab: add exact comparison twins and the MessagePipe semantic ledger #507 owns the MessagePipe SubUnsub twin and semantic ledger.
• perf lab: retain and replay content-addressed experiment evidence #508, test: build a differential bus state-machine oracle #509, and perf lab: calibrate variance, controls, power, and confirmatory inference #510 supply evidence, correctness, and decisions.
• perf lab: resolve IL2CPP cost with WPR, PMU, and native maps #511 supplies native evidence for claimed interface, atomic, or copy costs.

Exploration may begin earlier. Production confirmation waits for the #500 gates.

Purpose

Reduce registration and deregistration time, allocation calls, copied state, and retained storage without moving steady dispatch or weakening stale-handle safety.

Current representation

MessageRegistrationToken already has a slot array, doubly linked live list, and free list. This issue must not recreate that mechanism under a new name.

Concrete attributed candidates include:

• one process-wide atomic ID increment per staged registration;
• large copied MessageBusRegistration and typed deregistration state;
• Dictionary plus List ordered spill storage with a dictionary lookup during ordered iteration and O(n) arbitrary list removal.

Owned experiments

• Attribute every registration/deregistration allocation and copied byte by owner.
• Measure direct bus, direct handler, token stage/active, remove, and disable at 1/16/256/1,000 live registrations.
• Test owner-local generational handles: one token-owner identity plus slot and generation, with a burned slot on wrap.
• Keep the internal handle at or below the current 16 bytes and preserve the public factory path where process-wide uniqueness is required.
• Test a bus-owned generational registration-record arena with a small opaque built-in handle while preserving custom IMessageBus ExternalId/ExternalState behavior.
• Test a compact ordered spill store: sparse hash index to append-ordered dense records, scrubbed tombstones, and transactional density-based compaction.
• Apply the ordered-store experiment to HandlerActionCache spill first. Extend it to bus handler/priority stores only after that mechanism wins.
• Test indexed or intrusive removal separately from storage layout.
• Test bounded slabs or pools only for an attributed allocation owner.
• Prototype batch register, enable, disable, and deregister transactions that validate, invalidate, and roll back once.
• Measure delegate equality/hash cost for static, instance, capturing, equal, and duplicate delegates.
• Revisit exact-MessageBus deregistration only when perf lab: resolve IL2CPP cost with WPR, PMU, and native maps #511 still shows the relevant indirect call and perf lab: calibrate variance, controls, power, and confirmatory inference #510 makes the lifecycle bracket interpretable.
• Measure cold prewarm, high-water capacity, occupied slots, idle sweep, force trim, and retained Unity references.

Do not repeat object-only token recycling or rejected inline-plus-ordinary-spill layouts without a materially different mechanism.

RED and correctness

• Extend test: build a differential bus state-machine oracle #509 with copied, double, foreign-token, out-of-order, stale, reset, reuse, and forced-wrap handles.
• Cover owner destruction, partial failure, remove/register during dispatch, duplicate refcounts, and remove/re-add ordering.
• Use LeakWatcher around 10,000 complete cycles.
• Return occupied slots and retained buffers to configured caps after sweep or trim.
• Clear every delegate, captured object, and Unity object reference before reuse.
• Keep every emit allocation row at zero.
• Preserve custom IMessageBus round trips and retryable failure behavior.

Acceptance

• The declared lifecycle endpoint passes perf lab: calibrate variance, controls, power, and confirmatory inference #510; the normal target boundary remains greater than 3%.
• Registration-cycle GC.Alloc calls do not increase. An allocation candidate must reduce its declared count or establish a structural zero.
• Handle size, copied bytes, probes, tombstones, compactions, and retained capacity are reported.
• Dispatch affected rows and equivalent sentinels pass their registered gates.
• Differential, lifecycle, leak, reclamation, allocation, and public-surface gates pass.
• A sealed holdout supports the result.
• Every representation receives a retained decision.

Route maps and middleware are out of scope.

[wallstop]

Adjacent structural result from #573 and PR #580:

Ordinary MessageAwareComponent subclasses now avoid exactly three default retained registrations. Explicit string-demo opt-in still retains the same three handlers. The allocation matrix passed 90 tests locally, and the corrective head d8bd8ac57f6a5113605c17a08401eda3efdd354e passes all CI workflows, including Performance Numbers.

This is a structural retained-state reduction discovered during the adoption walkthrough. It is not completion or attribution of #502's registered subscription-churn experiments. No #502 checklist item changes.