test: investigate intermittent full-stack dispatch allocation failure

[wallstop]

Failure

A native Unity 6000.4.6f1 EditMode run reported a managed allocation in AllocationMatrixTests.EmitWithFullStackIsZeroAlloc, labeled EmitFullStack-Untargeted.

The run selected both Editor and Editor.Allocations assemblies through the local MCP bridge. It completed with 1,219 passes, one failure, one skip, and eight inconclusive imported-sample cases in 277.987 seconds. The sample cases are a separate local fixture limitation. The raw result and passive framework completion status are preserved in the session 266 local artifacts.

Expected behavior

The warmed bus dispatches through an allowing interceptor, priority handlers, and postprocessors without managed allocations. The zero-allocation assertion must remain meaningful and reliable.

Investigation so far

• This test constructs a plain MessageBus and MessageHandler. It does not exercise MessagingComponent or the current owner-lifecycle fixes.
• Idle eviction is disabled. Before the fixture fix, diagnostics inherited the caller's global target; they were not guaranteed disabled. The emit closure is built before measurement. The helper warms both the action and the measurement wrapper.
• All 50 isolated native repetitions passed. The original 75-case allocation fixture also passed. A repeat of the combined assembly scope passed its allocation row and retained eight imported-sample inconclusives. These results do not explain or dismiss the original failure.
• The installed host Unity Test Framework 1.6.0 constraint was inspected. It calls Recorder.FilterToCurrentThread(), matching the relevant CI 1.4.5 behavior. Ordinary background allocation is not an established explanation. The inspected source is preserved locally in .artifacts/session266/installed-framework-source.txt.
• Current ordinary and scheduled Unity CI exclude the Allocation category. Historical benchmark artifacts have expired, so a green historical workflow is not exact test evidence.

Confirmed fixture isolation defect

Implemented in PR #551 at commit 0a7b798ff44ae009f4235004cfe34905132e7094.

A controlled native probe calls the actual fixture setup and full-stack row. Before the fix, input DiagnosticsTarget.All remains All after setup. All three canonical kinds fail with 193 GC allocation calls across 32 measured iterations. The bus and token both report diagnostics enabled. Input Off passes all three kinds. Stack traces are disabled in this experiment.

Fixture setup now selects Off and restores caller settings in teardown. The explicit diagnostic budget rows still select All inside their test bodies. The same post-fix native probe passes all six cases: both input All and input Off become Off after setup. The probe confirms restoration of all five altered globals.

Six state regression cases repeat default-Off and explicit-All checks across the three canonical kinds. They inspect setup, bus/token construction, and restoration; they do not measure dispatch. The existing allocation matrix covers dispatch. The post-fix full allocation assembly passes 291 cases with zero failures, skips, or inconclusives in 266.4863964 seconds. Run 304297c5-bccd-4da2-8b5f-831ff95ae500 has verified cleanup and restoration of the original clean scene.

The before/after records and probe are local, ignored artifacts under .artifacts/session268/allocation-inherited-diagnostics-{before.json,after.json,probe.cs}. They are not durable downloadable evidence. This experiment proves a fixture defect and a failure mechanism. The original failure did not capture diagnostics settings, so its historical cause remains unproven.

Remaining work

1. Preserve the completed isolated results and continue capturing every outcome under the original combined scope.
2. Repeat the original combined suite scope with diagnostics state captured to test whether order matters.
3. Capture state and allocation call stacks at a reproduced failure; identify the responsible production, test, or framework path.
4. Attribute and fix any remaining original failure without retries, relaxed thresholds, skipped tests, or artificial delays.
5. Repeat focused and full-scope verification and record the evidence.

A separate source review found duplicate postprocessor registration in the full-stack scenario. Correcting that coverage gap must not be presented as a proven fix for this allocation failure.

[wallstop]

The 50 isolated native runs are complete: 50 passed, zero failed/skipped/inconclusive. Each run reached the separate passive framework-cleanup marker. The assertion, warmup, and measured iteration counts were unchanged.

This establishes a 0/50 isolated reproduction rate in this editor session. It does not explain the original full-scope failure. Next checks are the installed constraint source and broader allocation-fixture ordering.

[wallstop]

The unchanged AllocationMatrixTests fixture also passed all 75 cases (35.261 seconds, zero failed/skipped/inconclusive). The installed host is Unity Test Framework 1.6.0; direct source inspection confirms FilterToCurrentThread() around the action and CollectFromAllThreads() after recording stops.

The next repetition preserves the original combined Editor and Editor.Allocations selection. A passive callback now retains the full final result tree and timestamps once at RunFinished, so reasons and ordering can be audited without active MCP polling.

[wallstop]

The original combined Editor + Editor.Allocations scope completed again with 1,220 passes and zero assertion failures in 254.779 seconds. Complete result capture confirms eight inconclusive sample-fixture cases and one sample-guide skip; these are the known missing imported samples.

EmitWithFullStackIsZeroAlloc passed. Its immediate predecessors were the diagnostics-enabled Broadcast, TargetedWithoutTargeting, and BroadcastWithoutSource rows, which also passed. No test or allocation threshold changed between the original failure and this repetition.

The original allocation failure is still unexplained. The confirmed postprocessor-priority coverage defect is tracked separately in #546.

[wallstop]

#547 includes the independent full-stack priority coverage correction from #546. On that branch, ten corrected-row runs and all 277 allocation-assembly cases pass with unchanged allocation thresholds. This issue remains open for attribution of the original intermittent failure; no root cause has been established.

[wallstop]

PR #551 adds failure-only bus state capture and expands the full-stack row across all five message kinds, including both without-context paths. Failure text retains the original allocation count and adds emission ID, token enablement, global and effective bus/token diagnostics flags, stack-capture configuration, and all six registration counters. Context is rendered once after the recorder stops; the measured action is not retried. Contract cases verify both failure reporting and omission on success.

Current local Unity 6000.4.6f1 evidence: unchanged baseline fixture 75/75 passed; final full allocation assembly 285/285 passed, zero failed/skipped/inconclusive, in 261.235 seconds. All five full-stack rows and the reporter contracts pass. The maintained runner reached passive cleanup with matching GUID/path and restored the original scene. Current installed constraint source again confirms FilterToCurrentThread around the action.

The original failure did not recur. No allocation call stack or attributed root cause was obtained, so this issue remains open. The change improves scope and the evidence available at a future failure; it does not claim to fix that unexplained allocation. PR CI is pending.

[wallstop]

Captured-state combined-scope non-recurrence on 2026-09-11 at 3107ea8719a226d92e57881115741058b49c5ac6:

• Unity 6000.4.6f1 EditMode selected WallstopStudios.DxMessaging.Tests.Editor plus WallstopStudios.DxMessaging.Tests.Editor.Allocations.
• The explicit documentation-image writer was excluded to avoid modifying user-owned generated assets. No category, allocation threshold, retry, measurement attempt, timeout, or production/test code changed.
• Result: 1,233 passed / 0 failed / 1 existing skip / 8 known imported-sample inconclusives, 300.1244842s.
• All five EmitWithFullStackIsZeroAlloc cases passed, including Untargeted.
• Run GUID: e0d70796-e186-4e12-ad49-c628679c319c.
• The complete raw result and matching run/cleanup records are retained locally under ignored .artifacts/unity-mcp/session287-editor-allocation-combined-545.json*.
• Cleanup had no observer/framework errors, found the framework inactive, and restored the original clean saved scene. The cleanup status intentionally fails closed because the eight sample nodes are inconclusive. A final passive query found all current/owned/legacy result-path keys empty.

The fixture already captures global diagnostics target, stack-trace setting, bus/token diagnostic modes, emission ID, and registration counts if the allocation recurs. It did not recur here, so this run provides no allocation call stack and does not attribute or dismiss the historical failure.

The exact Assets/DxmCiSamples fixture still cannot be copied safely into this shared host: its copied GUIDs collide with existing imported sample assets. That limitation accounts for the same eight inconclusives and was not worked around by renaming or mutating user assets.