perf: calibrate the IL2CPP measurement system and attribute dispatch cost

[wallstop]

Part of #497. This issue coordinates the measurement foundation for #414 and #501 through #505. It does not ship runtime optimizations.

Purpose

Create one reproducible measurement, correctness, evidence, and attribution system for the campaign. Detailed implementation belongs to the bounded tasks below.

Foundation tasks

• perf lab: freeze the canonical IL2CPP profile and add a shipping-fidelity player #506 freezes the canonical IL2CPP profile and adds a stripped shipping-fidelity player.
• perf lab: add exact comparison twins and the MessagePipe semantic ledger #507 adds exact comparison twins and the MessagePipe semantic ledger.
• perf lab: retain and replay content-addressed experiment evidence #508 retains and replays content-addressed evidence through immutable GitHub prereleases.
• test: build a differential bus state-machine oracle #509 builds the differential bus state-machine oracle.
• perf lab: calibrate variance, controls, power, and confirmatory inference #510 calibrates variance, controls, power, and confirmatory inference.
• perf lab: resolve IL2CPP cost with WPR, PMU, and native maps #511 resolves IL2CPP cost with WPR, PMU, and managed-to-native maps.
• perf lab: define cold, burst, and queued tail-latency protocols #512 defines cold, open-loop burst, queued, and tail-latency protocols.
• perf lab: validate ARM64, thermal, and energy behavior #513 validates oldest-Unity, ARM64, thermal, and energy behavior.

Start #506, #508, and #509 in parallel. #507 consumes the frozen profile and evidence format for timed conclusions. #510 consumes #506 through #508. #511 consumes #506 through #508 and #510. #512 consumes #506, #508 through #510. #513 consumes #506, #508, #510, and #512 and runs after canonical x64 confirmation.

Gates supplied to runtime workstreams

Confirmation gate

Before a production candidate from #501 through #504 enters confirmation:

• the canonical profile is machine-checked;
• the evidence bundle is content-addressed, replayable, and independently restored;
• the differential oracle covers the affected state transitions;
• the independent sample unit, repetition count, controls, intervals, multiplicity, and stopping rule are fixed by perf lab: calibrate variance, controls, power, and confirmatory inference #510;
• discovery and sealed holdout traces are frozen.

The existing 3% reducer remains a legacy screen until #510 approves confirmatory inference.

Mechanism gate

A claim that a named IL2CPP operation was removed requires replicated native attribution, a bound on all unresolved samples, profiler observer-effect evidence, and before/after native output from #511. Injected delays show sensitivity and dose response, not a maximum possible speedup.

Comparison gate

A #414 parity claim requires the exact twins and semantic ledger from #507. Unlike mutation or lifecycle semantics must remain explicit.

Latency and platform gates

• Latency-sensitive and queued claims use perf lab: define cold, burst, and queued tail-latency protocols #512.
• Cross-platform default promotion uses perf lab: validate ARM64, thermal, and energy behavior #513.
• ARM64 characterization does not block early Windows x64 screening, but an architecture-sensitive default must pass it or remain conditional.

Campaign housekeeping

• Synchronize the campaign decision ledger with the fix: reject null token handlers before registration mutation #420 correctness result, the accepted perf: reduce registration churn and harden Unity MCP readiness #432 detached spare, later rejected Research MessagePipe performance targets and formal parity for 4.0.0 #414 candidates, and the measurement-blocked exact-MessageBus specialization.
• Link every accepted, rejected, inconclusive, invalidated, and blocked experiment to its immutable evidence bundle.
• Keep GitHub as the durable source of truth. Keep PLAN.md below 100 lines as a transient routing page; detailed protocols belong in runbooks, and local session records belong in ignored progress/.

Completion

This issue is complete when every foundation task is complete or closed with a retained reason, the decision ledger is current, and the resulting protocols are linked from the relevant workstreams. Failed profiling paths, negative controls, and measurement blockers are completion evidence.

No runtime optimization ships from this issue.

Durable evidence index

The tracked evidence index and restore record is included in PR #571. Repository-wide immutability is enabled with maintainer approval. The first historical Mono/debug observation is published as an immutable prerelease and restored from a fresh remote clone; full archive/manifest hashes and exact replay match. The real-asset denied-access drill returns HTTP 401 without a cached fallback. #521 remains open for index review and landing. This storage result does not satisfy the remaining measurement, IL2CPP, or confirmatory campaign gates.

[wallstop]

PR #524 synchronizes the campaign decision ledger in 331aecc with reviewed GitHub sources. It distinguishes accepted correctness/allocation changes, rejected dispatch experiments, and measurement-blocked variants. It also records later native code-generation progress without treating tool inventory as capture evidence or historical throughput changes as causal proof.

The ledger links the actual decision records and preserves missing immutable raw evidence as a gap. Independent source review found no remaining discrepancies. This is documentation of known outcomes, not completion of the #506/#508/#509/#510 foundation gates.

[wallstop]

The versioned MessagePipe semantic ledger is now directly linked here for the parity/foundation audit:

• Ledger: https://github.com/Ambiguous-Interactive/DxMessaging/blob/1590dd334982cd832e65401138b5bd11904d520b/scripts/unity/comparison-semantic-ledger-v1.json
• Evidence catalog: https://github.com/Ambiguous-Interactive/DxMessaging/blob/1590dd334982cd832e65401138b5bd11904d520b/scripts/unity/comparison-evidence-catalog-v1.json
• Schema: https://github.com/Ambiguous-Interactive/DxMessaging/blob/1590dd334982cd832e65401138b5bd11904d520b/scripts/unity/comparison-contract-v1.schema.json

The ledger covers all nine public rows and 108 required dimension entries. It remains draft and parityEligible: false; linking it does not establish parity or finish native semantic evidence.

[wallstop]

The completed #506 protocol is now linked from the two required routing surfaces at source 3107ea8719a226d92e57881115741058b49c5ac6:

• Canonical verdict profile
• Build and runtime configuration protocol, including the distinct Minimal/Low/Medium/High shipping-fidelity profiles, clean-build provenance, AOT/stripping topology matrix, and evidence limits

These are versioned protocol links, not a claim that native player payload retention is complete. #506's content-addressed build/result evidence item remains open.

[wallstop]

PR #591 CI evidence at fc5ee446:

• Hosted script CI is green on Linux, macOS, and Windows. The three test-suite durations total 315.11s versus 325.40s on the prior successful PR head: -10.29s (-3.2%) aggregate. Per-runner movement (+6.7s Windows, +2.5s Linux, -19.5s macOS) is treated as hosted-runner noise, not a platform speed claim.
• Unity correctness is green on 2021.3.45f1, 2022.3.45f1, and 6000.3.16f1.
• The 6000.5.2f1 internal IL2CPP job passed all 202 outcomes. Its unchanged artifact-redaction gate completed in 3.073s, redacted the retained Bee DAG, and uploaded artifact 10330773769, archive SHA-256 beeb9fe1536e55d73168a189bf120193d8df8b424b0704b86064bfcc8520e5a0.
• License return, cleanup classification, and organization-lock release passed.

This establishes the formerly blocked internal native upload at the reviewed source. It does not yet establish final acceptance: the 6000.5 correctness leg and comparison IL2CPP job are still queued, and human review remains required.

[wallstop]

Foundation dependency rollup (2026-09-14)

Current evidence changes the next safe action but does not complete this coordinator:

• perf lab: add exact comparison twins and the MessagePipe semantic ledger #507 and test: build a differential bus state-machine oracle #509 are complete.
• perf lab: freeze the canonical IL2CPP profile and add a shipping-fidelity player #506 has its canonical and shipping-fidelity mechanisms, but content-addressed build/result retention remains open.
• perf lab: retain and replay content-addressed experiment evidence #508 has four registered text-evidence classes and a verified immutable restore path. PR Index immutable endpoint evidence #592 indexes the two shipping endpoint bundles, but still awaits wallstop review; native payload publication remains fail-closed.
• perf lab: calibrate variance, controls, power, and confirmatory inference #510 has a detailed candidate preregistration, but the independent unit resolves to a three-build palindrome. The resulting 72-binary floor (96 at the invalid-block cap), health characterization, and target-only control require wallstop approval before implementation. The legacy bracket remains screening-only.
• perf lab: resolve IL2CPP cost with WPR, PMU, and native maps #511 now has a readiness audit. Selected generated-C++/PDB/source/disassembly plumbing is verified, but no WPR/ETL, PMU, workload-interval, observer-effect, sampled-share, or unresolved-share experiment exists.
• perf lab: define cold, burst, and queued tail-latency protocols #512 now has a readiness audit. Existing cold rows are same-process type-cold diagnostics; fresh-player cold, p99/open-loop, and queue protocols remain absent.
• perf lab: validate ARM64, thermal, and energy behavior #513 now has a device readiness audit. Native Linux ARM64 devcontainer CI is not an ARM64 Unity player/device result; device-state, sustained thermal, and energy controls remain absent.
• Runtime screen ordering is recorded in perf: screen the simple untargeted dispatch hot path #501 and the perf: screen route-plan data structures and middleware pipelines #503–perf: prototype frozen, batched, queued, and Burst-compatible messaging tiers #505 readiness ledgers. None may be promoted through confirmation until perf lab: calibrate variance, controls, power, and confirmatory inference #510 is approved and completed; perf lab: resolve IL2CPP cost with WPR, PMU, and native maps #511 attribution is additionally required before a mechanism-specific native-cost claim.

Therefore the immediate coordinator path is: obtain human review for PR #592; resolve #510's explicit approval questions before spending self-hosted runner time; then implement the #511/#512 evidence contracts through #508. Keep all foundation boxes above unchecked unless their own issue-level acceptance evidence lands.

[wallstop]

Foundation rollup after PR #593

The predecessor-snapshot correction, payload ABI contracts, and target-map topology observers are now merged to master at b6ce2f59f9c3e2115708226d51d7586c3e8846d1. All 39 PR checks passed.

#506 endpoint run 34901087633 is queued at that exact source with the manual-only clean/incremental shipping factor. Its dispatch superseded the ordinary same-ref Unity push run through the existing concurrency group, so it does not duplicate the four-version correctness matrix. Recurring CI topology and coverage are unchanged.

#501 and #503 now have their screening contracts on the default branch, but neither can enter confirmation: #510's palindrome-unit/cost and control-design approval remains unresolved, and #511 still lacks its admitted trace contract. The next foundation decision after #506 evidence acceptance is therefore still #510 approval; no confirmatory runner campaign should start before that decision.

[wallstop]

Foundation rollup update: the clean/incremental endpoint factor for #506 is now complete through immutable publication and independent restore.

• Exact merged-source validation: https://github.com/Ambiguous-Interactive/DxMessaging/actions/runs/34901087633
• Unity 2021.3 corrected immutable bundle: https://github.com/Ambiguous-Interactive/DxMessaging/releases/tag/perf-evidence-shipping-fidelity-matrix-2021.3.45f1-r2-474a82d0
• Unity 6000.5 corrected immutable bundle: https://github.com/Ambiguous-Interactive/DxMessaging/releases/tag/perf-evidence-shipping-fidelity-matrix-6000.5.2f1-r3-eb7d15ae
• Durable index review: Index clean and incremental shipping evidence #596

Both endpoint jobs passed 20 clean shipping cells plus their separately bound incremental build. Fresh anonymous restores verified and replayed both bundles. PR #596 is documentation-only; all 27 checks are complete (24 success, 3 expected skips), and its Unity gate used the docs-only path without a licensed matrix.

#506 should close only after human review and merge of the durable index. Native payload publication remains excluded by the privacy gate. This does not approve the separate #510 campaign spend or establish native attribution for #501/#503.

[wallstop]

Published immutable s336 Editor open-loop evidence, revision 1: release, tracked evidence index at commit 70a63207. The exact archive SHA-256 is a7bb9e40e8afe83e691a9c4a5c7154986fc94ba2aff8e6facca75ef1f2c13259; manifest SHA-256 is 7da590ac55e36a8a8a92c6e9e5c2fe3c325189e6f71eea0c9d28b1352b10fb8e; bundle digest is 8ef62f844f1c3e4533679df2df7ec19cc50955fcbe1ad8748425c3c6b4a2871f. Release 390857962 reports draft:false, prerelease:true, immutable:true; tag resolves to verifier commit 79279da3d50e52a7e452e86f5ef515a873b69be3, and asset 570580334 matches the archive digest. Draft and postpublication downloads independently matched; fresh remote verifier clones with pinned dependencies passed verify and exact replay. Anonymous postpublication download succeeded. An invalid-credential request for the actual asset returned HTTP 401 with no file or fallback. This is descriptive Unity Editor PlayMode Mono evidence from measured source fe8fe99ec0382eaa2b85ad054b51209776f1d53c, not an IL2CPP or promotion result.

The foundation now has one remotely restored exact-tree Editor open-loop capture. Its player and broader campaign gates remain open.

[wallstop]

The preregistered Editor clock retention candidate (protocol) passed. The strict editor-latency-clock-capture-v1 reducer at verifier commit f22d7ab35c214a23b3332b5ffcb67a74b30a6252 sealed both seven-file scrubbed captures, recomputed their exact raw-pair summaries, bound terminal runner sidecars and measured source commits, and rejected missing, altered, nonterminal, source-drifted, and mismatched replay inputs. Full local npm test passed 1,331 Node plus 27 and 56 Python tests; repository validation, format, spelling, and hooks passed. Static CI run 35246207564 completed success, including the reducer suites on Linux, macOS, and Windows.

Published immutable s337 single-pair controls (release 390880212, asset 570627932, archive SHA-256 f058f0ebc9fdafb2c8cf545f2055e31933be0b7b899d5ca3d4a6a8ccc9b48ec9, manifest SHA-256 9c890828c941e03e8e515bd9704b0254fec36452db0df3d9a973e106166053a5) and immutable s338 batched controls (release 390880898, asset 570629904, archive SHA-256 1aa797afbb89c77f3c3210e848918bbd79bef8935f30799d276a9b60e159b8ac, manifest SHA-256 6f6df9ec760fc4d11d15b3ec3d37719a44c36ece4a166bcbb38477a3aaeb2f80). Each draft's sole asset downloaded and replayed from a fresh remote verifier checkout before publication. After publication, both releases reported draft:false, prerelease:true, immutable:true; their tags resolve to the verifier commit. Separate anonymous downloads matched all indexed hashes, and a new clean remote checkout passed both verify and exact replay. Invalid credentials against each actual asset API ID returned HTTP 401, curl exit 22, and no file or fallback. Tracked evidence index at commit 84cff56b records identities and restore limits.

The result is durable descriptive Editor Mono clock evidence. s337 remains quantized and s338's elapsed windows include loop/checksum work. Neither establishes individual player p99 resolution, IL2CPP cost, or a latency promotion verdict. The branch awaits PR/human review when ELI's exact license reservation is safe.

These controls add durable, source-bound measurement diagnostics to the foundation; they do not clear the player calibration dependency.