perf: screen the simple untargeted dispatch hot path

[wallstop]

Part of #497 and #414.

Inputs and dependencies

• perf: calibrate the IL2CPP measurement system and attribute dispatch cost #500 coordinates the measurement foundation.
• perf lab: freeze the canonical IL2CPP profile and add a shipping-fidelity player #506 supplies the canonical and shipping-fidelity players.
• perf lab: add exact comparison twins and the MessagePipe semantic ledger #507 supplies exact simple-dispatch twins.
• perf lab: retain and replay content-addressed experiment evidence #508 supplies retained evidence.
• test: build a differential bus state-machine oracle #509 supplies the differential state-machine oracle.
• perf lab: calibrate variance, controls, power, and confirmatory inference #510 supplies screening and confirmation rules.
• perf lab: resolve IL2CPP cost with WPR, PMU, and native maps #511 supplies native attribution and code maps.
• perf lab: define cold, burst, and queued tail-latency protocols #512 supplies cold and burst characterization.

Exploration may begin earlier. Production confirmation waits for #506, #508, #509, and #510. A named native-cost claim also waits for #511.

Purpose

Decompose and reduce fixed cost in the simple untargeted class and struct route without weakening the dynamic public contract.

Owned experiments

• Screen sweep-gate, plan lookup/validation, dispatch lease, emission/tick/touch stores, diagnostics/debug branches, live activity, reset polling, token callback augmentation, casts, callback ABI, and 0/1/N shape.
• Use a class control, the exact Research MessagePipe performance targets and formal parity for 4.0.0 #414 comparison struct, and readonly 0/16/64/256-byte structs at 0, 1, and 16 handlers.
• Use a secondary 1/2/4/8-machine-word ABI matrix only to locate calling-convention transitions.
• Compare by-value and readonly-in callbacks across monomorphic and heterogeneous targets.
• Test relevance-scoped plan epochs with hot type A plus unrelated type-B churn.
• Test guided lease code shape only if perf lab: resolve IL2CPP cost with WPR, PMU, and native maps #511 places an interval-supported material share in lease handling.
• Compare the current generic-static message index and route lookup with an explicit resolved type index, bus-bound prepared emitter, and generated closed bus-local leaf.
• Stop a resolved-type or generated candidate when native output shows no class-init, RGCTX, cast, cache, or bounds cost to remove.
• Compare the current lazy first-post-mutation rebuild with eager immutable publication for the simple scalar route. perf: screen route-plan data structures and middleware pipelines #503 owns keyed and middleware extension.
• Test selected no-feature kernels, diagnostics mode selection, typed cached routes, hot/cold outlining, and evidence-guided entry layout.
• Test manual prefetch or blocked traversal only at fan-out 64/256 and only after PMU evidence.

WPR capture, PMU collection, evidence retention, power calculation, and comparison-bridge implementation belong to the linked lab tasks.

RED and correctness

• Add route-specific operations and observations to test: build a differential bus state-machine oracle #509.
• Cover ordering, found results, diagnostics, exceptions, live disable/re-enable, reset during dispatch, same/cross-kind reentrancy, nested throw, deferred reset, stale prepared handles, trim, and multiple buses.
• Prove every mutation-to-epoch mapping before scoping invalidation.
• Prove eager publication retains the old route for the current emission and exposes the new route to the contractually correct nested or later emission.
• Keep every affected allocation-matrix row at zero.
• Preserve live MessageHandler.active and the per-callback reset-generation reread unless an additive frozen contract explicitly removes mutation.

Acceptance

Every screened factor receives an effect estimate and a keep, park, reject, inconclusive, invalid, or measurement-blocked decision.

A production candidate must:

• change the intended native mechanism;
• pass perf lab: calibrate variance, controls, power, and confirmatory inference #510, with the normal target boundary remaining greater than 3%;
• preserve affected rows and equivalent sentinels;
• pass differential, lifecycle, reentrancy, exception, diagnostics, allocation, leak, reclamation, and public-surface gates;
• pass a sealed holdout;
• record native symbols, player size, and build time;
• pass applicable perf lab: define cold, burst, and queued tail-latency protocols #512 and perf lab: validate ARM64, thermal, and energy behavior #513 gates before broad claims.

Registration storage, target maps, middleware pipelines, and additive semantic modes are out of scope. Do not repeat a rejected #414 representation without a new causal reason.

[wallstop]

Preregistration: untargeted idle-sweep gate attribution

Question. Does the steady-state TrySweepIdle call account for a material share of simple untargeted class and struct dispatch, before any actual sweep occurs?

Baseline. Every typed untargeted emission calls TrySweepIdle. With idle eviction enabled, 15 of 16 calls return after the emission-mask check; the sixteenth also reads the clock before the configured interval check. The current retained comparison harness can pair DxMessaging with MessagePipe in one process.

Diagnostic ablation. Temporarily remove only the TrySweepIdle() call from UntargetedBroadcast<TMessage>. Do not change sweep implementation, settings, timing windows, dispatch state, or tests. Compare current/candidate/restored-current builds in the same warm Unity editor. Run the existing paired GlobalToOne and StructNoBox rows, retaining all cycle ratios and the DxMessaging/MessagePipe ratio. MessagePipe is the causal host-movement sentinel.

RED. Source and loaded IL must contain the call in both current arms and omit it only in the candidate arm. The tracked file must return byte-for-byte to its starting hash after the experiment.

GREEN and decision. Each paired test must pass exact fan-out with no failures or skips. Treat a raw cycle spread above 3% or inconsistent class/struct movement as uninterpretable. A consistent ratio improvement strictly above 3% identifies the gate as a material target for a semantics-preserving amortization design; otherwise reject further gate work. This screen cannot authorize disabling reclamation or make a confirmatory claim.

Stop rule. Stop and restore immediately on a compile failure, dirty scene, framework cleanup failure, allocation/correctness failure, or any change outside the one temporary call-site deletion. Do not push the ablation or change recurring CI.

Evidence. Issue #501 and ignored .artifacts/unity-mcp/session321-* result, cleanup, identity, source-hash, and extracted paired-summary files.

[wallstop]

Idle-sweep gate attribution result: material local target, ablation restored

Preregistration: #501 (comment)

At source 8305a6759273de73add1591a96d0e7aecf014c7d, I ran the existing paired DxMessaging/MessagePipe rows in Unity 6000.4.6f1 PlayMode Mono x64 Debug. The sequence was current / temporary one-line ablation / restored current. The ablation removed only TrySweepIdle() from UntargetedBroadcast<TMessage>.

Fresh loaded-assembly inspection confirmed both current arms were 767 IL bytes and called MessageBus.TrySweepIdle. The candidate was 759 IL bytes and had no sweep call.

Scenario	Current A ratio	Candidate ratio	Current B ratio	Candidate effect vs control geometric mean	Outer control spread	Maximum raw-cycle spread
GlobalToOne	0.174756848	0.185634939	0.175498019	+6.000%	0.424%	1.468%
StructNoBox	0.197590082	0.205636591	0.196295728	+4.415%	0.659%	0.802%

All six scenario executions passed exact fan-out with zero failures, skips, or inconclusive results. Every owned cleanup record reports the framework inactive, PlayMode stopped, the main stage, and the original single saved clean scene.

Result SHA-256 values:

• current A: e165e7b1e811e9eaca2106cc311b00d34f3d30d206bab6e5b1fb9f4bfa7aa720
• candidate: 627fb2e85ccecaccb48f800226d902aa1f0a3f98f81e872c43ba76f626cc860d
• current B: 7bfa13a02f2b7b74c0f956b15e9f22ccff382bca742dd294fd8714da4db7dcd4

Decision: material attribution target, not a production candidate. The local screen clears the preregistered 3% threshold on both class and struct routes with stable MessagePipe controls. Do not remove the gate: that would disable on-emit reclamation and does not preserve the contract. Any follow-up must preserve low-throughput aging, non-PlayMode use, custom PlayerLoop replacement, explicit trim, and sweep-before-plan-validation ordering. It still requires #509 correctness, IL2CPP evidence, #510 confirmation, and retained #508 evidence.

The production file is restored byte-for-byte to SHA-256 21447a78fa67d1e64a946f09a38be412aca313444828baf47edeaa9be965dc37. The loaded method again contains the call. The tracked worktree has no runtime diff, and no test, workflow, timeout, or recurring CI work changed.

[wallstop]

Preregistration: outline the sampled idle-sweep slow path

Question. Can the current idle-reclamation contract keep its exact emit and PlayerLoop cadence while removing cold clock/sweep code from the simple dispatch hot body?

Causal basis. The retained one-line ablation above improved the local DxMessaging/MessagePipe ratio by 6.00% for GlobalToOne and 4.41% for StructNoBox, with stable controls. It proved the gate material in Mono but cannot be retained because it disables on-emit reclamation.

Candidate. Preserve the current enabled check and 16-emission mask in an aggressively inlined emit wrapper. Call a no-inline sampled slow method only when both pass. The slow method performs the unchanged clock interval check, optional idle-aging tick, and Sweep(false). Give the PlayerLoop an explicit wrapper that preserves its current enabled check and advanceTickForIdleAging: true behavior. Apply the same emit wrapper to untargeted, targeted, and sourced broadcasts. Do not change constants, settings, clock source, sweep order, plan invalidation, or public APIs.

RED. The current loaded generic untargeted method calls the monolithic TrySweepIdle; the slow clock/sweep body is not structurally isolated. The candidate must show the cheap wrapper in the generic method and a distinct no-inline slow method. A source diff outside the sweep refactor is a hard failure.

GREEN.

• Run current/candidate/restored-current paired GlobalToOne and StructNoBox rows. All raw-cycle and outer-control spreads must remain within 3%.
• Require a consistent candidate ratio improvement strictly above 3% on both rows before retaining the representation for IL2CPP confirmation.
• Pass the complete EmitGateClockReadIsRare fixture and EvictionSweepContractTests, including no-emit PlayerLoop aging, the first clock sample on emission 16, disabled eviction, and real sweep behavior.
• Keep affected emit allocation rows at measured zero and leave the public surface unchanged.

Stop rule. Restore and reject on any cadence, clock-read, sweep, PlayerLoop, cleanup, allocation, compile, or public-surface failure; also reject if either paired target misses 3% or controls are unstable. Even a local win is parked, not shipped, until #508/#510 and IL2CPP confirmation are available. Do not push the temporary candidate or change recurring CI.

Evidence. Issue #501 and ignored .artifacts/unity-mcp/session321-outline-* artifacts plus the local session record.

[wallstop]

Sampled slow-path outlining result: rejected and restored

Preregistration: #501 (comment)

The temporary candidate kept the enabled check and 16-emission mask in a 45-byte aggressively inlined wrapper. It moved the unchanged clock/interval/tick/sweep body into a 59-byte no-inline method. The PlayerLoop retained its current enabled check and idle-aging request.

Fresh loaded IL showed the generic untargeted caller calling only the eligibility wrapper. The complete EmitGateClockReadIsRare and EvictionSweepContractTests fixtures passed 46/46. The complete allocation matrix passed 85/85. Every owned cleanup record was clean.

The performance sequence used a current control immediately before the candidate and a freshly compiled restored current control after it:

Scenario	Current A ratio	Candidate ratio	Current B ratio	Candidate effect vs control geometric mean	Outer control spread	Maximum raw-cycle spread
GlobalToOne	0.175498019	0.176980771	0.176008358	+0.699%	0.291%	1.284%
StructNoBox	0.196295728	0.203011798	0.196976940	+3.242%	0.347%	0.686%

Verdict: reject. GlobalToOne improved only 0.699%, so the candidate failed the required consistent improvement strictly above 3%. The stable controls and raw cycles make this an interpretable negative result. The struct-only movement does not permit a route-specific keep under the preregistration.

Evidence hashes:

• correctness result: 962547fa3ef8b7cc891acd3d90e0efb4c1b6eef631222be29aad80ccb7b25f74
• allocation result: ecef9a6ad743588d385b66fc837999198f039002ce8d7c3692ea7afe97667b23
• candidate paired result: f96f640a16aa2c5a5c834b303cab375f4df86ea28c44decef0a913913cdd3803
• restored-current paired result: edb43e8b1822e89b0f2ce23b963fc970504ecb7635fbad0371e8829da1b2f35f

The production file is restored byte-for-byte to 21447a78fa67d1e64a946f09a38be412aca313444828baf47edeaa9be965dc37. The loaded method is again the original 767-byte caller, the original 107-byte aggressive-inline gate is present, and the candidate methods are absent. No tracked candidate, test, workflow, timeout, or recurring CI change remains.

[wallstop]

Preregistration: untargeted diagnostics-off branch attribution

Question. Does the live _diagnosticsMode check and its cold recording block account for a material share of simple untargeted class and struct dispatch when diagnostics are off?

Baseline. The paired benchmark asserts both bus and token diagnostics are off, but every UntargetedBroadcast<TMessage> still evaluates _diagnosticsMode before the fast path. The preceding restored-current run is a valid immediate outer control.

Diagnostic ablation. Temporarily delete only the if (_diagnosticsMode) { _emissionBuffer.Add(...) } block from typed untargeted dispatch. Do not change diagnostics defaults, other routes, timing windows, fan-out, or tests. Run restored-current / candidate / freshly restored-current paired GlobalToOne and StructNoBox rows with MessagePipe as the causal host-movement sentinel.

RED and GREEN. Fresh loaded IL must contain the diagnostics field read in both current arms and omit it only in the candidate. Both paired cases must pass exact fan-out. Raw-cycle and outer-control spreads must stay within 3%. A consistent candidate ratio improvement strictly above 3% on both rows identifies diagnostics mode selection as a material design target; otherwise reject further work on this branch.

Stop rule. This ablation intentionally breaks diagnostics-on recording and can never ship. Restore immediately after the candidate run. Stop on compile, fan-out, framework-cleanup, scene, or control-stability failure. Do not push it, change recurring CI, or infer an IL2CPP result.

Evidence. Issue #501 and ignored .artifacts/unity-mcp/session321-diagnostics-* artifacts plus the local session record.

[wallstop]

Diagnostics-off branch attribution result (follow-up to the preregistration):

The temporary candidate deleted only the _diagnosticsMode recording block from typed untargeted dispatch. Fresh loaded IL shrank from 767 to 724 bytes and no longer referenced the diagnostic field; after restoration it returned to 767 bytes and referenced _diagnosticsMode.

Scenario	Current A	Candidate	Current B	Candidate vs control geometric mean	Outer control spread	Max raw-cycle spread
GlobalToOne	0.176008358	0.176619121	0.178334436	-0.310%	1.322%	0.878%
StructNoBox	0.196976940	0.197278139	0.197636900	-0.014%	0.335%	0.544%

Both candidate cases passed. Neither result approached the preregistered >3% threshold, and the controls were stable. Decision: reject further work on the diagnostics-off branch and keep the live diagnostics behavior.

Candidate source SHA-256: 37cfec3d16843780a79c3314602c9d9df6bacfa296e0e6b571cd55986b7c37f7.

Candidate result/run/cleanup SHA-256:

• 13b01c3515cf271f2ac56fb7ac4477d328010fe2a57329780d6b5d7c510e0218
• 502dbac515fd1fb869f44e4b67d31ae0b1e1f13069016071e552df8960f6469b
• 086bfb0f01866e91780275f916232f6b29410bcba8b124c8e5b26f87552e6aa3

Restored-current-B result/run/cleanup SHA-256:

• 70c2a1d78bda8bcd5d919d7530c27923a81c2b831b8d703a139391728bdee8ef
• 37ff7ee5869dbc7b88abf79e3f1d35558a05e4bb2b4dd9258a9a3ba3b8528380
• fa85828af4df2f1c54d3b151a3a150eebf2e575cfedda63821b5c28645c3867a

Production source was restored byte-for-byte (SHA-256 21447a78fa67d1e64a946f09a38be412aca313444828baf47edeaa9be965dc37). No runtime, test, workflow, timeout, or recurring-CI change remains.

[wallstop]

Preregistration: untargeted emission-tick increment attribution

Question. Does the per-emission AdvanceTick() increment account for a material share of simple untargeted class and struct dispatch, separately from the already-rejected lastTouchTicks store candidate?

Baseline. Typed untargeted dispatch increments _tickCounter before entering its plan-specific path. The fast cached one-handler path then writes the resulting tick to the handler cache. The previous freshly restored run (session321-diagnostics-control-d.json) is the immediate current-A control.

Diagnostic ablation. Temporarily replace only long touchTick = AdvanceTick(); in UntargetedBroadcast<TMessage> with long touchTick = _tickCounter;. This retains the local value and downstream touch stores while removing only the increment/call at this route. Do not change the emission ID, sweep gate, diagnostics, plan, lease, handler loop, timing windows, or tests. Run candidate and freshly restored-current GlobalToOne and StructNoBox pairs, with MessagePipe as the host-movement sentinel.

RED and GREEN. Fresh loaded IL must call AdvanceTick in both current arms and omit that call only in the candidate. Both paired cases must pass exact fan-out. Raw-cycle and outer-control spreads must stay within 3%. A consistent ratio improvement strictly above 3% on both rows identifies tick advancement as a material design target; otherwise reject further work on this factor.

Stop rule. The ablation intentionally stops emission-driven idle aging and can never ship. Restore immediately after the candidate. Stop on compile, fan-out, cleanup, scene, or control-stability failure. Do not push it, change recurring CI, or infer native/IL2CPP impact.

Evidence. Issue #501 and ignored .artifacts/unity-mcp/session321-tick-* artifacts plus the local session record.

[wallstop]

Emission-tick increment attribution result: rejected and restored

Preregistration: #501 (comment)

The temporary candidate replaced only long touchTick = AdvanceTick(); in typed untargeted dispatch with long touchTick = _tickCounter;. It retained the local value and downstream touch stores while intentionally disabling the increment for attribution. Fresh loaded IL stayed at 767 bytes but contained zero AdvanceTick calls; restored current contained one call again.

Scenario	Current A	Candidate	Current B	Candidate vs control geometric mean	Outer control spread	Max raw-cycle spread
GlobalToOne	0.178334436	0.179090517	0.174419966	+1.545%	2.244%	0.788%
StructNoBox	0.197636900	0.201172894	0.196167443	+2.170%	0.749%	0.544%

Both candidate and restored-current cases passed with clean runner cleanup. All stability bounds passed, but both effects missed the required strictly-greater-than-3% gate. Decision: reject further work on the emission-tick increment and preserve emission-driven idle aging.

Candidate source SHA-256: 5529e75ddbb5f4b4abc6a918b540c41838f02e8c9fb71bc8a14092ae665f3e65.

Candidate result/run/cleanup SHA-256:

• ecfff8232d3ddd18a53afa30babb4f89534d1b6ce4869d7347a4ec37b672e994
• 72b4e5150092a90b5940dae12c089af2a579b26be014a08e73074788a9b17cff
• 6fa8c298577cad8ef0034de06f43f67ed3d32c372e7eb3dd4afa05fe74383ae5

Restored-current-B result/run/cleanup SHA-256:

• 8a6fd0de604d0178d456f22877c34f16450d4438c86fc577aab1cd121d8ef6fd
• 173a1e124e914a00085628992569fb51af2d670525efbea809e6179312a325d7
• 36bf9d300877f33c66be0456df2c70b57b6741699d0b8a2996b0d32a1bcb5fd7

Production source is restored byte-for-byte to 21447a78fa67d1e64a946f09a38be412aca313444828baf47edeaa9be965dc37. No runtime, test, workflow, timeout, or recurring-CI change remains.

[wallstop]

Preregistration: untargeted plan-version validation attribution

Question. Does the warmed plan.version != _dispatchPlanVersion validation and cold refresh edge account for a material share of simple untargeted class and struct dispatch, separately from the generic plan lookup?

Prior evidence. The campaign ledger rejects moving a cached-route check ahead of the handler-count check, but it does not contain an ablation of the plan-version validation itself. Session 056 listed this factor for a local pre-filter and session 057 left it unmeasured. The freshly restored session321-tick-control-e.json run is current A.

Diagnostic ablation. In UntargetedBroadcast<TMessage>, temporarily replace the current version read and refresh conditional with long planVersion = plan.version;. Keep the generic _untargetedDispatchPlans.TryGetValue<TMessage> lookup, plan object, sweep gate, lease, emission/tick/diagnostics operations, handler loop, timing windows, and tests unchanged. Run candidate and freshly restored-current paired GlobalToOne and StructNoBox rows, with MessagePipe as the host-movement sentinel.

RED and GREEN. Fresh loaded IL must reference _dispatchPlanVersion and call RefreshUntargetedDispatchPlan in both current arms, and omit both from the candidate's generic method. Both cases must pass exact fan-out. Raw-cycle and outer-control spreads must remain within 3%. A consistent candidate ratio improvement strictly above 3% on both rows identifies validation as a material design target; otherwise reject further work on this isolated factor.

Stop rule. This ablation intentionally fails to refresh stale plans after registration mutation and can never ship. Restore immediately after the candidate. Stop on compile, fan-out, cleanup, scene, or stability failure. Do not push it, change recurring CI, claim correctness beyond the mutation-free benchmark, or infer native/IL2CPP impact.

Evidence. Issue #501 and ignored .artifacts/unity-mcp/session322-plan-validation-* artifacts plus a local session record.

[wallstop]

Plan-validation screen attempt: invalid and restored

Preregistration: #501 (comment)

Fresh loaded IL showed that the scoped candidate removed RefreshUntargetedDispatchPlan, but the generic method still contained one _dispatchPlanVersion read in its later non-fast mutation fallback. The preregistration incorrectly required zero field references, so the candidate could not satisfy its stated RED criterion.

The already-started owned runner was allowed to finish only for safe cleanup (2/2 passed; framework inactive; original scene clean). Its timing is invalid and will not be used. Result/run/cleanup SHA-256:

• 1632da078afe1bed2b907b2496d1116c783f96119deb5b2844b16a4919b5181c
• 0a00ca85c97155477ca931f43ca382c469454b96dad42cba2bf3e37788c36de1
• 102915ee6b273d95c2bc0aeaed5605cde2989e9926710c1c261e019ce938c054

Production source is restored byte-for-byte to 21447a78fa67d1e64a946f09a38be412aca313444828baf47edeaa9be965dc37. A corrected preregistration will use resolved call/reference counts: current has two version-field reads and one refresh call; the scoped candidate must have one later fallback read and zero refresh calls.

[wallstop]

Corrected preregistration: untargeted plan-version validation attribution

This supersedes the invalid RED criterion in #501 (comment). No timing from that invalid arm is eligible.

Question and candidate. Isolate the warmed plan-version validation while retaining the generic plan lookup and later mutation fallback. Temporarily replace only the initial long planVersion = _dispatchPlanVersion; plus if (plan.version != planVersion) RefreshUntargetedDispatchPlan<TMessage>(plan); with long planVersion = plan.version;. Keep every other component unchanged. This remains a non-shippable mutation-semantics ablation.

Corrected RED. Resolve IL operands, including generic method-spec tokens. Current must be 767 bytes with exactly two _dispatchPlanVersion field reads and one resolved RefreshUntargetedDispatchPlan call. Candidate must be 739 bytes with exactly one version-field read (the intentionally retained later fallback) and zero refresh calls. Restored current must return to the current counts and exact source hash.

Measurement and GREEN. Use freshly restored session321-tick-control-e.json as current A, then run a new candidate artifact and a newly compiled current B for paired GlobalToOne and StructNoBox. Both cases must pass exact fan-out. Raw-cycle and outer-control spreads must remain within 3%. A consistent candidate ratio improvement strictly above 3% on both rows identifies the isolated validation as material; otherwise reject it.

Stop rule. Restore immediately after the candidate. Stop on any IL-count, compile, fan-out, cleanup, scene, or stability failure. Do not push, change recurring CI, reuse the invalid timing, claim mutation correctness, or infer native/IL2CPP impact.

Evidence. Issue #501 and new ignored .artifacts/unity-mcp/session322-plan-validation-valid-* artifacts plus the local session record.

[wallstop]

Corrected plan-validation attempt: invalid mechanism, restored

Corrected preregistration: #501 (comment)

The candidate met the corrected loaded-IL counts (739 bytes, one retained later fallback version read, zero refresh calls) and passed 2/2 cases with clean cleanup. However, removing the only refresh call also meant a newly created plan was never populated. Warmup therefore remained on the slow path instead of isolating the warmed version-validation branch.

The observed ratios are invalid mechanism output, not effect estimates, and no current-B run was taken:

• GlobalToOne: 0.128181410
• StructNoBox: 0.146005484

Result/run/cleanup SHA-256:

• 26d82f2846463f38dc0e42e33631c93bcb17c42efbbba0cea297ebcf8b34b9c3
• e671eb87c55ab84f2a6d7a507db4563a03492798b0390da0d161d34f9831dde6
• 88e738129af0f7aea3b7b7886d125e094e53ec65169a1ed9ab664147d4c9bd22

Production source is restored byte-for-byte to 21447a78fa67d1e64a946f09a38be412aca313444828baf47edeaa9be965dc37. A valid screen must retain first-plan refresh and remove only subsequent warmed validation.

[wallstop]

Final preregistration: warmed plan validation with first-touch refresh preserved

This supersedes both invalid attempts above. Their timings remain ineligible.

Question. Does only the repeated warmed plan.version != _dispatchPlanVersion validation materially affect simple untargeted dispatch?

Current and candidate. Take a new current A first. The candidate keeps RefreshUntargetedDispatchPlan<TMessage>(plan) on the first plan-creation branch, immediately after the existing AOT bridge call, then assigns long planVersion = plan.version; without the repeated validation conditional. The comparison fixture registers before its first warmup emit, so first creation observes the registration and publishes the fast route. Keep the generic plan lookup, later mutation fallback, and all other dispatch work unchanged. The candidate remains non-shippable because a later registration mutation would not refresh an existing plan.

Mechanism gates.

• Current loaded IL: two resolved _dispatchPlanVersion reads and one resolved refresh call.
• Candidate: one retained later-fallback version read and one retained first-creation refresh call.
• A reflection probe after one candidate emit must find fastPath == true, one handle entry, and a populated cached handle route for both the class and exact comparison struct.
• Restored current must return to its counts and exact source hash.

Measurement. Run new current-A / candidate / newly compiled current-B paired GlobalToOne and StructNoBox artifacts. Both cases must pass exact fan-out; raw-cycle and outer-control spreads must remain within 3%. A consistent candidate ratio improvement strictly above 3% on both rows identifies the warmed validation as material; otherwise reject it.

Stop rule. Stop and restore on any mechanism, route-state, compile, fan-out, cleanup, scene, or stability failure. Do not push, alter recurring CI, use either invalid timing, claim post-mutation correctness, or infer native/IL2CPP impact.

Evidence. Issue #501 and new ignored .artifacts/unity-mcp/session322-plan-validation-final-* artifacts plus the local session record.

[wallstop]

Warmed plan-validation attempt: invalid across sweep boundary, restored

Final preregistration: #501 (comment)

New current A passed 2/2:

• GlobalToOne 0.174422755, raw spread 1.225%
• StructNoBox 0.196911512, raw spread 2.457%

The candidate preserved first-plan refresh, reduced loaded IL from 767 to 747 bytes, retained one resolved refresh call, and reduced _dispatchPlanVersion reads from two to one. A post-first-emit reflection probe confirmed both class and struct plans were fastPath == true, with one populated cached handle entry.

The candidate then split sharply after its first measurement cycle:

Scenario	Headline	Raw cycle ratios	Raw spread
GlobalToOne	0.141225173	0.177767635, 0.130715121, 0.130573849, 0.131103335	36.143%
StructNoBox	0.156359336	0.197118091, 0.147551835, 0.143655120, 0.143055006	37.792%

The first cycles matched current; periodic idle sweeping then invalidated the plan. Without repeated validation, later cycles retained a stale plan and moved to the slow route. Both spreads exceed the fixed 3% gate by an order of magnitude. Decision: invalid as an isolated timing ablation; no effect estimate and no current-B run.

This retains a useful constraint: first-touch refresh alone cannot remove warmed validation across a real window. Any production design must preserve immediate registration invalidation and sweep-driven refresh, such as the separately owned relevance-scoped epoch investigation.

Current-A result/run/cleanup SHA-256:

• 3887b0546d6dfb8ff062370d0805631c96f0ab60a2fca30f154bc39f19460d95
• 5d98daef11805b92fda16621f06c339b2081bbbb047f8afa8142313a7bfa3cc9
• 12d0bb7c81c6196ef9b21084c4fb5dd3595cccc5c0dc62e989988e35061e5f6f

Candidate result/run/cleanup SHA-256:

• d33c8d958ddfca5a17b08c738ecdf76d41fe28dc268a1d57e2031562728392d0
• 4f1e619b0c99723840ca020547cc553028861c0a7d32477d570b340735ada1c6
• 63c597c577267f9792716cbf38b54229e1390db3a2fc03b80e0890b273de96da

Production source is restored byte-for-byte to 21447a78fa67d1e64a946f09a38be412aca313444828baf47edeaa9be965dc37; loaded current is again 767 bytes with two version reads and one resolved refresh call. No runtime, test, workflow, timeout, or recurring-CI change remains.