perf: screen route-plan data structures and middleware pipelines

[wallstop]

Part of #497 and the filtered, post-processing, and keyed outcomes in #414. Consumes the shared-prologue baseline from #501.

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 exact keyed and middleware comparison twins.
• 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 attribution and code maps.
• perf lab: define cold, burst, and queued tail-latency protocols #512 supplies burst characterization for accepted candidates.

Purpose

Measure and optimize targeted, broadcast, priority, global-accept-all, interceptor, post-processing, and rewritten-route costs without conflating them with the simple emit prologue.

Owned experiments

• Add exact one-hook, one-post, one-plus-one, 16-keys-one-hit, one-token/four-priority, global-accept-all, targeted, broadcast, and without-context topology rows.
• Run an orthogonal covering screen over hit/miss, cardinality, handlers, active density, priorities, hook depth, veto/rewrite rate, and churn.
• Profile current IntKeyMap probes, load, collisions, resize, delete moves, hit/miss, key skew, cache events, and retained bytes.
• Add sanitized Unity keys, contiguous IDs, negative/extreme IDs, power-of-two strides, aligned low bits, uniform random, Zipf access, mixer-specific collisions, and misses inside/outside clusters.
• Screen hash mixer, load factor, probing, deletion, and layout as mechanisms instead of treating named maps as indivisible candidates.
• Compare dynamic managed candidates: current linear/backward-shift storage, Robin Hood, scalar tagged-metadata open addressing, and standard Dictionary controls.
• Use Swiss or F14 terminology only when native evidence proves vector control-byte group comparison. Faithful SIMD variants belong to perf: prototype frozen, batched, queued, and Burst-compatible messaging tiers #505.
• Keep sorted vector, dense offset/page, Eytzinger, and perfect-hash candidates in frozen route work handed to perf: prototype frozen, batched, queued, and Burst-compatible messaging tiers #505.
• Do not repeat inline 1/2/4 target maps plus ordinary spill; prior physical inline context maps failed retained storage.
• Measure priority containers during registration, arbitrary removal, and snapshot build. Test a bitset only for an explicitly bounded priority domain with a general int fallback.
• Test 0/1/N hook kernels, fused dynamic programs, post snapshot cost, route rewrite, global flattening, and relevance-scoped invalidation.
• Compare AoS/SoA, packing, and numeric width only after profile/probe evidence identifies a removable cost.

RED and correctness

• Extend test: build a differential bus state-machine oracle #509 with all affected route, hook, rewrite, priority, and context operations.
• Audit every mutation site against its required epoch.
• Preserve first-registration and priority order, emission-start snapshots, live activity, veto/rewrite, reset, reentrancy, and next-emission hook visibility.
• Cover key 0, signed extremes, collision clusters, cluster deletion, resize failure, stale route handles, and full managed-reference scrubbing.
• Keep dispatch allocation, leak, reclamation, and public-surface contracts green.

Acceptance

Every route family and data-structure mechanism receives baseline, probe/profile, allocation, retained-memory, and decision evidence.

A production winner must:

• pass perf lab: calibrate variance, controls, power, and confirmatory inference #510, with the normal target boundary remaining greater than 3%;
• win on a preregistered observed workload stratum rather than only an aggregate mean;
• preserve registered p99 probe and memory limits;
• preserve affected rows and equivalent sentinels;
• pass independent IL2CPP confirmation, a sealed holdout, and every correctness gate;
• record symbol, code-size, player-size, and build-time effects.

Shared prologue, compiler-only work, additive frozen semantics, and queue architecture are out of scope.

[wallstop]

Historical factor provenance and current gap audit

Audit source: 8305a6759273de73add1591a96d0e7aecf014c7d (2026-09-14). This maps pre-#503 work into this issue so accepted and failed mechanisms are not silently repeated. It does not upgrade historical screens to #510 confirmation.

Retained production foundations

Mechanism	Current status	Evidence boundary
Four active per-context maps changed from Dictionary<InstanceId,T> to the current integer-key, backward-shift IntKeyMap<T>	keep / current baseline	PR #340 retained exact zero-allocation rows, materially lower retained construction bytes, smaller IL2CPP player, and improved integrated targeted/broadcast rows against two runtime-identical controls. The controls also exposed up to 18.45% row noise, so isolated timing rows remain diagnostic rather than modern #510 confirmation.
Interceptors flattened to copy-on-write priority-ordered arrays owned by the interceptor store; dispatch plan caches the store	keep / current baseline	PR #412 removed scratch-list/group copying, fixed cross-priority mutation freezing, and added first-emit-after-unrelated-churn allocation coverage. The ownership correction is essential: keying the flat array to the bus-wide plan stamp allocated 4,096 bytes over 128 churn-interleaved emits; store-local dirty state measured zero.
Post routes frozen at emission start with keyed historical snapshots for rewrite/mutation cases	keep / correctness baseline	PR #415 fixed the empty-sink live-lookup inconsistency and added nested mutation, reclamation, and allocation coverage. It is the semantic floor for any fused post kernel.
Exact stable/rewritten targeted and broadcast post-route rows	keep / benchmark baseline	PR #423 retained six rows plus rewrite/fan-out contracts and generated-code capture after rejecting its runtime candidate.
Borrowed settled untargeted handle array/count	keep, but shared-prologue input from #501	PR #437 retained the current no-feature route with allocation, mutation, reset, trim, sweep, and reclamation guards. Do not count it as a new #503 middleware result.

Already-screened candidates

Factor / representation	Decision	Reason
Targeted post-phase seven-argument bundle replaced by one readonly by-ref state	reject	The exact A/B/B/A IL2CPP rows regressed targeted stable/rewrite shapes by 16.22–32.21% while broadcast sibling controls stayed within 3%; generated C++ retained pointer/getter loads. PR #423.
Bounded 256+ scalar open-addressing wrapper around the current target map	reject	256-key hit improved only 2.83%, below the screen; the wrapper added an object to dominant small maps and had remove-time allocation/throw, incomplete managed-key clearing, deleted-load growth, incomparable retention, and non-transactional migration defects.
Physical inline context maps with capacities 2, 4, and 8 plus ordinary spill	reject; do not repeat	All failed the unconditional 5% spill-retention boundary: capacity 2 increased four-key bytes 13.3% and slots 28.6%; capacity 4 increased sixteen-key bytes 9.8% and slots 23.5%; capacity 8 retained 47.1% more slots and over 3× measured bytes at sixteen keys.
Embedded inline priority owners with capacities 2, 4, and 8 plus ordinary Dictionary/List spill	reject; do not repeat	All retained extra embedded storage after spill while keeping the full legacy collections; each failed the 5% retained-storage gate before any timing promotion.
AggressiveInlining on InterceptorCache.EnsureFlat	inconclusive, reverted	Native output removed the intended call, but untouched GlobalToOne and StructNoBox moved about +7.3%; the corrected bracket rejected the causal claim. #414 correction.
Load cached handle array before the handler dictionary-count gate	inconclusive, reverted	Correct native shape and green semantics, but three unreachable IL2CPP sentinels moved beyond ±3%. Evidence.
Borrow settled post handle/post arrays into the untargeted plan	inconclusive, reverted	Green mutation/allocation/reclamation gates, but three unreachable sentinels moved beyond ±3%; the apparent +29.3% PostProcess effect is not causal evidence. Evidence.
Borrow interceptor flat array/count into the dispatch plan	inconclusive, reverted; do not rescreen unchanged	Both candidate binaries removed the exact native EnsureFlat call and passed behavior/allocation/reclamation gates. Four unreachable sentinels moved beyond ±3%, so the apparent Filtered effect was uninterpretable. Evidence.
Thread-local snapshot-holder pool	reject	Added per-thread retained state over the process-wide pool ceiling; failed the explicit memory gate before a bracketed timing claim.

What the current benchmark tree actually covers

• ComparisonTopologyBenchmarks has exact internal twins for keyed 16-keys/one-hit, four priorities on one token, one interceptor, one post-processor, one-plus-one, churn, and the canonical struct. These are real five-second windows with exact accounting; do not duplicate them under new names.
• DispatchThroughputBenchmarks adds targeted miss/one/16 listeners, one sourced-broadcast listener, four interceptors, four post-processors, and stable/rewritten-empty/rewritten-populated post routes for targeted and broadcast.
• TargetMapBenchmarks covers hit/miss/churn and physical topology at 1/4/16/256/4,096 keys, but its timed keys are only contiguous positive IDs in three separated ranges.
• IntKeyMapTests covers key 0, signed extremes, wraparound collision deletion, and high-bit mixing for correctness. Those tests are not performance evidence for sanitized Unity IDs, negative/extreme distributions, strides/aligned low bits, random, Zipf, or designed collision clusters.
• No dedicated performance row was found for global-accept-all, targeted-without-targeting, or broadcast-without-source. No covering array, cache-event profile, probe-count distribution, or retained-byte result exists for the requested key families.

Next non-duplicative slice

On a separate branch after the active documentation PR lands:

1. Add only the missing global-accept-all, targeted-without-targeting, and broadcast-without-source topology rows, with fast/action overload identity, exact fan-out, diagnostics-off, teardown, and allocation contracts. Reuse the existing one-hook/one-post/one-plus-one twins instead of cloning them.
2. Parameterize target-map key generation independently of operation/cardinality. First add contract-only deterministic generators for sequential sanitized IDs, signed extremes, power-of-two strides/aligned low bits, uniform seeded random, designed current-mixer collisions, and hit/miss probes inside/outside clusters.
3. Preregister a bounded covering set before timing it; do not form the full Cartesian product. Keep existing 1/4/16/256/4,096 cardinality and construction rows as common controls.
4. Treat every historical inconclusive candidate as unavailable without a materially new representation and a new perf lab: calibrate variance, controls, power, and confirmatory inference #510-compliant bracket.

Still unscreened

Robin Hood and tagged-metadata maps, load-factor/probing/deletion factors, global flattening, 0/1/N hook kernels, fused dynamic middleware programs, relevance-scoped invalidation, key-skew/cache-event attribution, and AoS/SoA/packing/width changes remain unscreened. SIMD terminology, frozen route layouts, and manual prefetch remain gated exactly as the issue states.

No #503 checklist item is marked complete by this provenance audit: the issue requires current baseline/probe/allocation/retention evidence and #510/#511 gates. This comment only supplies authoritative factor decisions and prevents duplicate work.

[wallstop]

Missing-topology slice: local candidate verified

Prepared the preregistered first missing slice locally on codex/issue-503-topology-rows at commit 2a569633 (not pushed and no stacked PR while #592 awaits human review):

• TargetedFlood_OneWithoutTargetingHandler — one token-owned TargetedWithoutTargeting registration, fast context overload, diagnostics off, exact fanout 1.
• BroadcastFlood_OneWithoutSourceHandler — one token-owned BroadcastWithoutSource registration, fast context overload, diagnostics off, exact fanout 1.
• GlobalAcceptAll_UntargetedClass_OneHandler — one token-owned GlobalAcceptAll registration, class untargeted message, diagnostics off, exact fanout 1.

The harness now assigns stable scenario keys/display names and has exact topology contracts for bus bucket counts, token state, registration kind/type/priority, and absent context.

RED / correction

The first focused EditMode run passed 2/3 and correctly rejected the global topology assertion: the captured registration type is DxMessaging.Core.IMessage, not the initially expected DxMessaging.Core.Messages.IMessage. After correcting that expectation, the focused contract run passed 6/6.

The first PlayMode execution also exposed one 24-byte allocation in the global row because the class message was constructed inside EmitMany. Moving the reusable instance to benchmark setup removed it from the measurement window.

Final local Unity evidence

• EditMode topology + fanout contracts: 6 passed, 0 failed, 0 inconclusive, 0.880 s; cleanup GUID 8603d247-c07d-4f75-a5d1-18c8724c7b5b.
• PlayMode three-row batch: 3 passed, 0 failed, 0 inconclusive. Targeted-without-targeting measured 0 allocations / 0 bytes; broadcast-without-source measured 0 / 0.
• Final global row after setup correction: 1 passed, 0 failed, 0 inconclusive, 5.759 s, 0 allocations / 0 bytes; cleanup GUID e9b67a5e-d796-400a-9eb8-ccf80ba76451.
• Host: Unity 6000.4.6f1, Editor PlayMode Mono x64 Debug on macOS. Fresh-assembly timestamp checks passed before the final runs; all cleanup sidecars matched their owned result paths, framework was inactive, main stage was restored, and the saved scene remained clean.
• Commit hooks passed (CSharpier, method-name normalization, llms/skill harness checks, devcontainer/meta validation).

These are local screening rows, not a published baseline or confirmation result. Rebase/push/open the #503 PR only after #592 lands.

[wallstop]

Preregistration: complete the missing-route callback ABI matrix

Question. Do the three newly added missing-route topologies have explicit, separately executable Action and fast-delegate rows without changing route topology, fanout, teardown, diagnostics, or steady-state allocation behavior?

Baseline. Local commit 2a569633 adds one fast-delegate row each for targeted-without-targeting, broadcast-without-source, and global accept-all. The historical factor contract called for fast/Action overload identity, but the current local scenario names do not expose the fast arm and no Action sibling rows exist. Existing token diagnostics normalize both overloads to the same public registration type, so metadata alone cannot prove ABI coverage.

Hypothesis. Rename the unpublished local rows with an explicit Fast suffix, add one matched Action sibling per topology, and bind each registration through an explicitly typed delegate before calling the overload. Both arms should retain identical one-token, diagnostics-off, no-context topology and exact fanout 1. The Action and fast arms remain separate five-second windows.

RED. Before the change, scenario/protocol enumeration has only three ambiguous rows and no Action sibling keys. The completed contract must fail if any of the six scenarios is missing, routes through a different bucket/context, invokes other than once, or reports a steady-state allocation.

GREEN. Require formatter/hooks; exact topology and fanout contracts for all six rows; focused Unity EditMode execution; focused Unity PlayMode execution of all six rows with positive throughput, exact handler accounting, and zero measured managed allocations/bytes. Preserve cleanup ownership and a clean shared editor scene.

Stop rule. Keep this as benchmark coverage only. Do not claim an Action-vs-fast performance effect from local Editor measurements, do not modify production dispatch, do not duplicate existing hook/post twins, and do not publish or stack a PR before #592 lands.

Evidence location. Local branch codex/issue-503-topology-rows, ignored session record, and a follow-up issue comment with RED/GREEN results.

[wallstop]

Callback ABI matrix result: keep as benchmark coverage

Preregistration: #503 (comment)

The local candidate is now complete at 6c11468e on codex/issue-503-topology-rows. The three unpublished ambiguous scenario names were replaced by six explicit rows:

• targeted-without-targeting: Action and fast
• broadcast-without-source: Action and fast
• global accept-all over an untargeted class message: Action and fast

Each registration binds an explicitly typed delegate before overload resolution. Action and fast siblings retain the same one-token, diagnostics-off, no-context bus topology, exact fanout 1, and independent five-second measurement windows.

GREEN evidence

• Fresh-assembly proof resolved all six final enum symbols before testing.
• Focused EditMode fanout/topology scope: 25 passed, 0 failed, 0 skipped, 0 inconclusive, 3.458 s; cleanup GUID d8a2f97a-cf21-4a3f-b2ba-75dc7f69f816.
• Focused PlayMode six-row scope: 6 passed, 0 failed, 0 skipped, 0 inconclusive, 30.137 s; cleanup GUID 46653ba9-0944-409f-80b9-9af0394b1be9.
• Every row reported 0 managed allocations and 0 allocated bytes with positive throughput and an approximately five-second wall-clock window.
• Commit hooks passed, including CSharpier, line endings, llms/skill drift, and Unity metadata validation.
• Both accepted cleanup records matched their owned result paths, reported no framework errors, and restored an inactive main-stage editor with the saved scene clean.

The Unity host necessarily stamped local structured rows with the overlaid primary worktree HEAD (8305a675); the tested candidate bytes are committed separately at 6c11468e. After testing, the primary three files were restored to their exact pre-overlay SHA-256 values and Unity was refreshed; the candidate symbols are absent from the loaded primary assembly.

Verdict and boundary

Keep the six rows as missing benchmark/contract coverage. The local Editor measurements are not an Action-vs-fast performance claim, a published baseline, or #510 confirmation. No production runtime code changed. The branch remains local and no stacked PR will be opened before #592 lands.

[wallstop]

Preregistration: deterministic target-map key-family contracts

Question. Can the existing target-map harness represent key distribution independently from operation and cardinality before any new timing rows are selected?

Baseline. All current hit/miss/churn rows generate contiguous positive IDs from three fixed ranges. Correctness tests cover isolated integer extremes but do not provide deterministic benchmark inputs for sanitized sequential IDs, signed extremes, aligned power-of-two strides, seeded random IDs, or clusters that collide under the current IntKeyMap mixer. Misses also have no explicit inside-cluster versus outside-cluster identity.

Hypothesis. Add a deterministic key-set generator with independent KeyFamily and MissProbeKind axes. Preserve the existing 15 timed rows byte-for-byte in identity by defaulting them to sequential sanitized/outside-cluster inputs. Add contract-only production-routing tests for every family and both collision miss probe kinds; do not expand the timed matrix yet.

RED. The current TargetMapBenchmarkCase has only cardinality and operation, TargetMapState synthesizes ranges internally, and no API can reproduce or distinguish the requested key families or miss-probe relationships.

GREEN. Require exact repeatability, uniqueness and disjointness, family-specific shape assertions, current-mixer bucket assertions for designed clusters, unchanged legacy scenario keys/count, exact hit/miss/churn routing and cardinality through the production map, clean teardown, formatter/hooks, and focused Unity tests.

Stop rule. Do not time the Cartesian product, change IntKeyMap, infer cache behavior, or claim one distribution is faster. If deterministic collision construction cannot remain bounded and exact at the retained cardinalities, reject that generator rather than weakening its bucket contract. A separate preregistration must freeze the bounded covering set before timing.

Evidence location. Local codex/issue-503-topology-rows branch, ignored session record, and follow-up issue verdict.

[wallstop]

Deterministic key-family contracts: keep

Preregistration: #503 (comment)

Prepared local commit f554ff6b on codex/issue-503-topology-rows. TargetMapBenchmarkCase now carries key-family and miss-probe axes independently of operation/cardinality, while every existing timed row defaults to its prior sequential sanitized/outside-cluster identity. No new timed case was added.

The contract-only generator supplies four mutually disjoint streams (registered, replacement, inside-cluster miss, outside-cluster miss) for:

• sequential sanitized IDs;
• signed extremes;
• power-of-two strides with aligned low bits;
• fixed-seed xorshift32 uniform IDs;
• exact designed collisions under the current IntKeyMap mixer.

The designed-collision generator inverts the current bijective 32-bit mixer, so construction is deterministic O(n) even at 4,096 keys instead of relying on an unbounded bucket search. Contracts then feed every generated key back through production IntKeyMap<object>.Bucket and through real MessageBus hit, miss, churn, sweep, registration, and teardown paths.

RED / correction

The original harness exposed only cardinality and operation and synthesized three contiguous ranges internally. After the first implementation passed 38 tests, critical review found that its default outside-miss stream used 0x4D30_0000 rather than the legacy 0x4D20_0000 range. The semantic stream ordering was corrected and the exact registered (0x4D00_0000), replacement (0x4D10_0000), and outside-miss (0x4D20_0000) ranges are now asserted.

Final GREEN evidence

• Complete TargetMapBenchmarkContractTests fixture: 38 passed, 0 failed, 0 skipped, 0 inconclusive, 39.740 s.
• Cleanup GUID c4eedb4d-8326-4a8e-ba36-ff702bc80f76; result ownership matched, no framework errors, framework inactive, main stage restored, saved scene clean.
• Exact collision bucket contracts passed at cardinalities 1, 4, 16, 256, and 4,096.
• Every family passed deterministic repeat, uniqueness/disjointness, production hit/miss/churn routing, physical cardinality, capacity, and original/current target assertions.
• Existing 15 scenario keys and default axes remained exact.
• Commit hooks passed.

Verdict: keep as input/contract infrastructure for the later bounded covering set. This is not timing evidence, a map candidate, a cache claim, or #510 confirmation. The primary PR worktree and loaded Unity assembly were restored to 8305a675; the local key-family and callback symbols are absent there. No stacked PR was opened.

[wallstop]

Preregistration: bounded target-map covering set

Question. What bounded set can cover the new key-family factors without timing the 5-family × 5-cardinality × 3-operation Cartesian product or removing the current controls?

Baseline. TargetMapBenchmarkScenarios.All contains the 15 legacy sequential-sanitized rows (five cardinalities × hit/miss/churn). Commit f554ff6b adds deterministic input axes but deliberately adds no timed rows.

Frozen selection. Keep all 15 legacy rows as common controls. Add exactly 20 non-sequential screening cases: for family index i = 1..4 in enum order and cardinality index j = 0..4 over [1, 4, 16, 256, 4096], select operation index (i + j) mod 3 over [Hit, Miss, Churn]. This gives one row for every non-sequential family/cardinality pair and covers every family/operation pair. For DesignedMixerCollision, its 1-key Miss uses InsideCluster; its 256-key Miss uses OutsideCluster; all other selected cases retain OutsideCluster.

The resulting screening inventory is exactly 35 unique cases: 15 unchanged controls plus 20 additions. It is intentionally not a full interaction design and does not estimate higher-order effects.

RED. No frozen screening inventory exists today; only the 15 automatic legacy rows exist. A later author could cherry-pick favorable families/cardinalities or accidentally expand the full Cartesian product.

GREEN. Encode a separate immutable/read-only screening list and contract-test exact count, uniqueness, inclusion of every legacy row, one non-sequential row per family/cardinality pair, all family/operation pairs, and both collision miss-probe kinds. Keep the NUnit TargetMapBenchmark source bound to the legacy All list so recurring Unity CI duration does not change.

Stop rule. Do not execute or publish timing from this commit. Do not add automatic test cases, choose rows from outcomes, replace legacy controls, or claim pairwise statistical identification. Execution needs a separate manual-only protocol and #510-compatible analysis preregistration.

Evidence location. Local codex/issue-503-topology-rows branch, ignored session record, and follow-up issue verdict.

[wallstop]

Bounded covering set: frozen, not executed

Preregistration: #503 (comment)

Local commit f0eea960 on codex/issue-503-topology-rows encodes the frozen screen as a separate read-only TargetMapBenchmarkScenarios.Screening inventory:

• 15 unchanged sequential-sanitized legacy rows remain the common controls.
• 20 non-sequential rows follow the preregistered Latin rule: one row per four-family × five-cardinality pair, operation (familyIndex + cardinalityIndex) mod 3.
• Designed-collision Miss uses InsideCluster at one key and OutsideCluster at 256 keys.
• Total inventory: exactly 35 unique rows.

The automatic NUnit benchmark source remains bound to TargetMapBenchmarkScenarios.All, which still contains exactly the original 15 keys and default axes. This commit therefore adds no five-second windows to recurring Unity CI and executes no screening outcome.

GREEN evidence

• Focused frozen-inventory + legacy-inventory Unity scope: 2 passed, 0 failed, 0 skipped, 0 inconclusive, 0.485 s.
• Cleanup GUID 407b827a-42b4-40be-bf84-3f095280ca56; result ownership matched, no framework errors, framework inactive, main stage restored, saved scene clean.
• Contracts prove the 35-case ceiling, read-only storage, uniqueness, every legacy key retained, exactly 20 non-sequential cases, every family/cardinality pair, every family/operation pair, and both collision miss-probe kinds.
• Commit hooks passed.

Verdict: keep as the outcome-independent selection contract. Timing remains unexecuted and requires a separately reviewed manual-only execution/retention protocol plus #510-compatible analysis. The primary branch and Unity assembly were restored to 8305a675; no candidate key-family symbol remains loaded and no stacked PR was opened.

[wallstop]

Preregistration: manual diagnostic execution of the frozen target-map screen

Question. Does the frozen 35-case inventory expose a large distribution/probe-shape signal worth promoting to later IL2CPP and #510 work, while preserving exact allocation/cardinality contracts?

Baseline. Commit f0eea960 freezes 15 legacy controls plus 20 outcome-independent additions. None has been timed outside the legacy sequential rows. The local endpoint is Unity 6000.4.6f1 Editor PlayMode Mono on macOS; it is a screening endpoint, not the shipping IL2CPP endpoint.

Execution. At exact local candidate f0eea960, execute every TargetMapBenchmarkScenarios.Screening entry once in its frozen list order through existing TargetMapBenchmarks.RunScenario. Each row retains its existing warmup, one continuous approximately five-second timing window, separate allocation probe, exact invocation count, map entries, and capacity. Use a fresh TargetMapState per row. Retain all 35 rows in one ignored CSV with source/host/session metadata and SHA-256; do not drop or rerun a row based on its result.

Primary diagnostic response. Operations/second by row, with zero-allocation and exact-cardinality validity gates. Ratios may only be described as observed local diagnostics against matching legacy operation/cardinality rows; no uncertainty or causal effect is available from one ordered session.

RED / GREEN. RED is the absence of retained observations for the frozen non-sequential rows. GREEN requires exactly the 35 frozen unique keys, positive operations and throughput, approximately five-second windows, exact entries/capacity/invocations, zero measured steady-state managed allocations/bytes, and a clean restored editor. Any failed validity gate rejects the affected observation and stops interpretation.

Stop rule. Do not repeat, reorder, trim, or promote based on favorable timing. Do not modify production code, add recurring CI work, name a winning map representation, or treat Editor Mono as IL2CPP/#510 evidence. If a gross key-family signal appears, retain it and use it only to preregister the next attribution or shipping-endpoint question.

Evidence location. Ignored .artifacts/issue-503-target-map-screen-session330.csv, ignored session record, and a follow-up issue comment.

[wallstop]

Pre-observation correction to the session330 allocation gate

The preregistration above incorrectly applied a zero-allocation gate to every operation. Churn deliberately removes, sweeps, and re-registers a handler inside its operation, so it measures registration lifecycle allocation and is not a promised steady-state dispatch path.

Corrected before executing any screen row:

• Hit and Miss: require exactly 0 measured managed allocations and 0 allocated bytes.
• Churn: require measured (not Unmeasured) non-negative allocation count/bytes and retain the exact values; do not reject the observation merely because lifecycle work allocates.
• All operations still require positive throughput/window, exact invocation accounting, exact entries/cardinality, and sufficient capacity.

No timing result had been observed when this correction was recorded. The frozen row set, order, execution count, interpretation boundary, and all other stop rules remain unchanged.

[wallstop]

session330 diagnostic screen: invalid, no timing verdict

Preregistration: #503 (comment)
Pre-observation allocation correction: #503 (comment)

All 35 frozen rows returned from Unity, but the orchestration failed before retaining the planned CSV. It therefore produces no interpretable timing result and will not be rerun under the same protocol.

Terminal evidence

The last progress marker was 35/35 rows complete. Post-collection validation then threw on:

{"index":2,"key":"TargetMap_1_Churn","keyFamily":"SequentialSanitized","missProbeKind":"OutsideCluster","keyCount":1,"operation":"Churn","totalOperations":190000,"operationsPerSecond":36341.345382236264,"wallClockMs":5228.2049,"gcAllocations":120000,"gcAllocatedBytes":11000000,"targetMapEntries":1,"targetMapCapacity":4,"observedInvocations":210000}

The row itself has exact cardinality/accounting and measured lifecycle allocation. The failure came from an orchestration-only wallClockMs <= 5200 check. That numeric ceiling was not frozen in the issue and is invalid: the benchmark targets an approximately five-second continuous window, and batch granularity can carry the final batch beyond 5.2 seconds.

More importantly, the orchestrator buffered all responses in memory and wrote the CSV only after validating the complete set. Throwing after collection discarded the other 34 returned rows. They cannot be reconstructed from durable evidence, so neither the terminal row nor partial console state may support a distribution comparison.

Verdict

Invalid / retain / no timing interpretation. No family, operation, cardinality, allocation comparison, or promotion decision is made. The ignored failure record is .artifacts/issue-503-target-map-screen-session330-invalid.md.

Per the stop rule, session330 was not repeated, reordered, or trimmed. A future attempt requires a new preregistration with retention-first writes for each response, positive-finite (not outcome-capped) time validation, a fresh session identity, and explicit handling of a mid-run persistence failure.

The primary file returned to its exact pre-overlay SHA-256, Unity was refreshed to 8305a675, the candidate type is absent, and the saved scene remains clean.

[wallstop]

Preregistration: session331 retention-first recovery screen

Question. Can the exact frozen 35-row diagnostic screen be collected once with durable per-row retention, fixing session330’s orchestration failure without changing selection or interpreting the invalid attempt?

Baseline. session330 is invalid because responses were buffered until post-run validation and an unpreregistered 5.2-second ceiling threw before persistence. Only its terminal row survives; it supplies no comparison. The benchmark/input code at f0eea960 passed its independent contracts.

Execution. Use fresh session identity session331 at candidate f0eea960, Unity 6000.4.6f1 Editor PlayMode Mono macOS. Pre-create the ignored CSV with metadata/header. Execute all 35 frozen rows once in the same list order. For each row: (1) receive the raw result, (2) append it durably to the CSV, then (3) validate it. If receipt, persistence, or validation fails, stop immediately and retain the partial CSV; never re-execute an index in this session.

Validity. Require unique frozen key/index, positive finite total operations/throughput/wall time, exact entries/cardinality, capacity at least cardinality, zero invocations for Miss, positive exact internally asserted accounting for Hit/Churn, 0 allocations/bytes for Hit/Miss, and measured non-negative allocation count/bytes for Churn. Do not impose an upper wall-time bound. RunScenario’s existing assertions remain authoritative for exact per-operation accounting.

Retention. After all 35 persist and validate, record CSV SHA-256 in a sidecar. Retain all rows, including unfavorable results. The invalid session330 row is not pooled, averaged, used as a control, or compared.

Interpretation/stop rule. Report only single-session local diagnostic ratios against the matching sequential operation/cardinality row, clearly without intervals or causal/shipping claims. Do not rerun, reorder, trim, promote a production representation, add CI, or treat Editor Mono as #510/IL2CPP evidence. Any row/persistence failure leaves session331 partial and terminal.

Evidence location. Ignored .artifacts/issue-503-target-map-screen-session331.csv plus .sha256, ignored session record, and a follow-up issue verdict.

[wallstop]

session331 retention-first screen: valid local diagnostic

Preregistration: #503 (comment)
Invalid predecessor (excluded): #503 (comment)

All 35 frozen rows were received, appended durably before validation, and validated without re-executing an index. The ignored CSV is 5,483 bytes and SHA-256 042dd05c18ebb86a69e469d40494d3b178b9f683e0bcbb560fc0f166231feb61; its sidecar verifies from the artifact directory.

Validity summary:

• 35 rows and 35 unique frozen keys.
• Every row has positive operations, throughput, and wall time plus exact production assertions for invocation accounting.
• Entries equal cardinality and capacity is sufficient in every row.
• All Hit/Miss rows report 0 allocation calls and 0 allocated bytes.
• All Churn rows retain measured non-negative lifecycle allocation; none is relabeled as steady-state dispatch.
• The primary worktree and Unity editor were restored afterward; candidate types are absent and the saved scene is clean.

Single-session observations follow. vs sequential divides by the matching legacy operation/cardinality observation from this same ordered Editor session.

#	Scenario	ops/s	vs sequential	ms	alloc calls	alloc bytes	entries/cap
0	TargetMap_1_Hit	1412652.781	1.000000x	5004.768	0	0	1/4
1	TargetMap_1_Miss	1732322.177	1.000000x	5004.843	0	0	1/4
2	TargetMap_1_Churn	29116.211	1.000000x	5151.769	120000	11000000	1/4
3	TargetMap_4_Hit	1399889.415	1.000000x	5007.538	0	0	4/8
4	TargetMap_4_Miss	1902327.018	1.000000x	5004.397	0	0	4/8
5	TargetMap_4_Churn	38598.000	1.000000x	5181.616	120000	11000000	4/8
6	TargetMap_16_Hit	1348953.093	1.000000x	5011.294	0	0	16/32
7	TargetMap_16_Miss	2057924.556	1.000000x	5000.183	0	0	16/32
8	TargetMap_16_Churn	34039.596	1.000000x	5287.959	120000	11000000	16/32
9	TargetMap_256_Hit	1293785.516	1.000000x	5000.829	0	0	256/512
10	TargetMap_256_Miss	1918486.544	1.000000x	5003.944	0	0	256/512
11	TargetMap_256_Churn	35443.440	1.000000x	5642.793	120000	11000000	256/512
12	TargetMap_4096_Hit	952871.876	1.000000x	5005.920	0	0	4096/8192
13	TargetMap_4096_Miss	2138924.677	1.000000x	5002.514	0	0	4096/8192
14	TargetMap_4096_Churn	35240.511	1.000000x	5107.758	120000	11000000	4096/8192
15	TargetMap_SignedExtremes_OutsideCluster_1_Miss	2075125.127	1.197886x	5002.108	0	0	1/4
16	TargetMap_SignedExtremes_OutsideCluster_4_Churn	31626.926	0.819393x	5058.980	120000	11000000	4/8
17	TargetMap_SignedExtremes_OutsideCluster_16_Hit	1425526.896	1.056765x	5001.659	0	0	16/32
18	TargetMap_SignedExtremes_OutsideCluster_256_Miss	1497462.531	0.780544x	5001.795	0	0	256/512
19	TargetMap_SignedExtremes_OutsideCluster_4096_Churn	32507.034	0.922434x	5229.637	120000	11000000	4096/8192
20	TargetMap_PowerOfTwoStride_OutsideCluster_1_Churn	37406.833	1.284742x	5079.286	120000	11000000	1/4
21	TargetMap_PowerOfTwoStride_OutsideCluster_4_Hit	1174821.020	0.839224x	5005.018	0	0	4/8
22	TargetMap_PowerOfTwoStride_OutsideCluster_16_Miss	1759506.599	0.854991x	5001.402	0	0	16/32
23	TargetMap_PowerOfTwoStride_OutsideCluster_256_Churn	30357.779	0.856513x	5270.478	120000	11000000	256/512
24	TargetMap_PowerOfTwoStride_OutsideCluster_4096_Hit	1187228.966	1.245948x	5003.247	0	0	4096/8192
25	TargetMap_UniformSeededRandom_OutsideCluster_1_Hit	1358893.670	0.961945x	5004.071	0	0	1/4
26	TargetMap_UniformSeededRandom_OutsideCluster_4_Miss	2108563.731	1.108413x	5003.406	0	0	4/8
27	TargetMap_UniformSeededRandom_OutsideCluster_16_Churn	35382.121	1.039440x	5087.315	120000	11000000	16/32
28	TargetMap_UniformSeededRandom_OutsideCluster_256_Hit	1340079.933	1.035782x	5007.164	0	0	256/512
29	TargetMap_UniformSeededRandom_OutsideCluster_4096_Miss	1812763.297	0.847512x	5003.411	0	0	4096/8192
30	TargetMap_DesignedMixerCollision_InsideCluster_1_Miss	1984880.845	1.145792x	5002.819	0	0	1/4
31	TargetMap_DesignedMixerCollision_OutsideCluster_4_Churn	40882.716	1.059193x	5136.645	120291	11015218	4/8
32	TargetMap_DesignedMixerCollision_OutsideCluster_16_Hit	1235748.204	0.916079x	5001.019	0	0	16/32
33	TargetMap_DesignedMixerCollision_OutsideCluster_256_Miss	2116850.550	1.103396x	5002.715	0	0	256/512
34	TargetMap_DesignedMixerCollision_OutsideCluster_4096_Churn	2206.037	0.062599x	9066.032	120007	11000386	4096/8192

Diagnostic boundary and next question

The non-sequential ratios range from 0.062599x to 1.284742x. Most are unsupported one-run movements and receive no promotion. The preregistered 4,096-key designed-collision Churn row is the sole gross signal: 2,206.037 ops/s versus 35,240.511 for matching sequential Churn (0.062599x), with the final batch extending the continuous window to 9.066 seconds. This is consistent with—but does not prove—long backward-shift deletion work in a deliberately single-bucket cluster.

Verdict: diagnostic keep, no causal/performance claim. session331 is not an interval, independent build unit, shipping IL2CPP result, cache-event profile, or #510 evidence. It selects only the next question: add non-timed exact probe-distance/cluster-length/removal-shift observations for the current map, with no hot-path instrumentation, then decide whether a bounded IL2CPP attribution run is justified. No production representation is selected or changed.

[wallstop]

Preregistration: exact target-map topology attribution

Question. Does the sole gross session331 signal—the 4,096-key designed-current-mixer-collision Churn row at 0.062599x its sequential control—correspond to exact lookup, cluster, and backward-shift deletion work in the current production IntKeyMap?

Baseline / RED. Candidate f0eea960 retains the timing result and deterministic key construction, while production exposes only target-map count/capacity. No current observer can establish probe distance, cluster length, deletion scan length, or removal moves. Timing alone is not causal evidence.

Hypothesis. At 4,096 registered keys and capacity 8,192, the designed-collision stream places every registered/replacement key in bucket 0. The initial map should therefore have one 4,096-entry cluster; registered lookups should require 8,390,656 total probes (1 through 4,096); the outside-cluster miss should require one probe; and a complete in-order primary-to-replacement churn pass should make each real removal scan and shift exactly 4,095 occupied entries, totaling 16,773,120 removal shifts while preserving cardinality and the single cluster. The matching sequential row is a retained descriptive control, not a thresholded candidate.

Implementation boundary. Add read-only, benchmark-only topology observation beside IntKeyMap and route it through the existing internal MessageBus benchmark bridge. Do not add counters, branches, writes, timers, allocation probes, or instrumentation to TryGetValue, Set, Remove, or DeleteAt. Run observation outside timed paths against the actual targeted-handle map. Execute exactly one real churn pass per selected case so each pre-removal prediction is tied to the state production then mutates.

GREEN checks. Independent contracts must cover empty maps, hits, misses, wraparound clusters, predicted deletion scans/moves, cardinality preservation, and the exact designed-collision invariants above. The existing target-key, matrix, routing, allocation, teardown, and full affected benchmark-assembly scopes must remain green.

Stop rule / interpretation. If the exact designed-collision invariants or observer/actual-mutation consistency fail, retain the mismatch and make no attribution. If they pass, attribute only the exact managed-map work selected by this deterministic input. Do not claim that it explains all elapsed time, general workloads, IL2CPP behavior, #510 significance, or a production winner. Do not time, rerun session331, add recurring CI timing, or change a map representation in this experiment.

Evidence location. Local branch codex/issue-503-topology-rows, ignored Unity result/raw structural artifact and session record, then a retained follow-up verdict here.