Compile Cache Input Freshness

A DLL's completion time does not establish which source snapshot it compiled. The runtime compiler therefore compares a cached artifact's persisted producing input digest with the generated input for the candidate snapshot before it reuses the artifact. Identical inputs still reuse their existing DLL and producer digest. Missing, different or inconclusive input identity follows the normal compilation path.

This is the same MeshNodeCompilationService path used by automatic compilation and the NodeType's normal Compile action. It does not add a second compiler, disable caching, modify source timestamps, or alter registry adoption and release hand-over policy.

The failure

The previous disk-cache predicate required a complete artifact set and a DLL write time at least as recent as the maximum source/NodeType LastModified and the framework timestamp. That allows this ordinary ordering:

1. A compilation captures source snapshot A.
2. An edit produces snapshot B while that compilation is running.
3. The A DLL finishes writing after B's LastModified.
4. The next compilation accepts the A DLL for B because its write time is newer.

The false hit also returned the current B source stamps beside A's bytes and producing digest. CompiledSources could consequently claim the new source had compiled while the emitted methods still implemented A. Automatic dirty detection then had a misleading success record to compare.

The normal forced Compile request bypasses dirty/prebuilt short-circuits but still enters this disk-cache path. Forcing a request alone did not repair its freshness predicate.

One captured snapshot, one input comparison

GetAssemblyLocationWithLog retains its single captured Source/Test snapshot and timestamp-based candidate lookup. For a candidate with a persisted digest it calls the existing generated-input regeneration logic with that exact snapshot. The shared path applies the canonical source/test shaping, include resolution, configuration skeleton, NuGet directive handling and input digest. It does not independently rediscover source nodes for the comparison.

Only a matching digest permits the hit. The successful hit still uses the persisted digest when stamping the dependency record; the regenerated candidate digest is evidence for the comparison, never a replacement producer claim. See Producer Determinism of the Dependency Record.

An unresolved NodeType definition is also inconclusive. It cannot be substituted with an empty configuration to establish cache equality. A present definition whose Configuration is null remains a valid input, so unchanged types with no configuration still take warm cache hits. This cache guard leaves the inherited fresh-compilation behavior for an unresolved definition unchanged.

The include reader uses the existing RunAsSystem boundary for each cold read, including the fallback read started by an earlier response. It restores the subscribing flow and downstream callbacks to their caller identity. A long-lived Observable.Using impersonation scope would instead be disposed by the response flow and can leave the subscriber elevated.

The existing re-evaluation entry point keeps its signature, discovery path and inconclusive result semantics. Neither CompilationCacheService's load-context lifecycle nor the registry's late hand-over machinery changes here.

Executed regression

The original regression ran against unmodified core 6231c4da4c1baa5dbcc9e44b43020006583fcb2c on 2026-09-11. It uses the real IMeshNodeCompilationService.CompileAndGetConfigurations and its existing sourcesOverride fixture seam, then invokes methods from the emitted assembly under the real scan pin.

The test derives B's LastModified as one tick before the first actual DLL's recorded completion time. This fixes the interleaving without a sleep, mocked compiler, or file timestamp edit. The before result was:

Expected: 43|new-case;stampsCurrent=True
Actual:   42|original-case;stampsCurrent=True
Second compiler transcript: Cache hit

The source method and test-registration method both remained old while the returned source stamps claimed B. One case executed and failed; fixture teardown completed cleanly. That historical red predates the expanded case matrix and is not a claim that every later case was run red.

The final CompileCacheInputFreshnessTest checks Source-only, Test-only, combined Source/Test and NodeType configuration-only changes. Every case owns a distinct type path because the standard fixture shares disk cache across instances of a test class. The configuration case invokes the actual emitted HubConfiguration, and the source cases inspect both emitted methods and CompiledSources.

The focused selection passed 50/50, including those four cases and the existing CacheHitStampsTheSameRecordTest, ContentKeyReevaluationTest and GeneratedInputIdentityTest controls. The compiler test project built in Release with warnings as errors, with zero warnings and errors. The unchanged-input control requires a real first emit, a subsequent actual cache hit, the same artifact path and an identical producer dependency record.

The full compiler suite first ran with the machine's default culture: 842/845 passed. The three failures were unchanged ShortWriteIsNotAPublicationTest assertions that require comma grouping (65,536, 4,096, 16,384) while the local runtime formatted Swiss apostrophes. The test and the relevant numeric-message formatter files are byte-identical to the baseline. With native xUnit's explicit -culture en-US, those three controls passed 3/3, then the full compiler suite passed 845/845, with no errors, skips or unrun tests. No assertion or formatter was changed.

The CI workflow runs the native xUnit host on ubuntu-latest without a culture override. The explicit local en-US setting is a reproducible test environment, not a claim that a CI job's actual culture was observed. Both the default-culture failure receipt and the explicit-culture success receipt are retained.

Review boundary regressions

PR review identified two additional boundaries in the shared regeneration path. Four cases ran against the unchanged PR head 96c157e3: the present-definition/null-configuration warm-cache control passed, while the missing-definition case and both real mesh source-read cases failed. The missing definition still produced a cache hit. The direct reader and the include fallback both left the subscribing caller holding system-security after Subscribe returned.

The amended cases require a present null configuration to keep reusing the original artifact, but an unresolved definition to fall through to the existing fresh compiler. The real mesh read cases also assert the returned node and caller identity inside the result callback, so restoring an identity by preventing the read from working cannot pass. They invoke the existing private reader boundary through reflection without exposing a new production API. The amended focused selection passed 54/54 after a strict Release build with zero warnings or errors. The amended full compiler suite passed 849/849 with explicit en-US, zero errors, skips or unrun tests. The earlier default-culture failure and pre-review receipts above remain historical evidence.

Reproduce with the normal SDK fixture:

dotnet build test/MeshWeaver.Compiler.Pipeline.Test/MeshWeaver.Compiler.Pipeline.Test.csproj -c Release -warnaserror
dotnet test test/MeshWeaver.Compiler.Pipeline.Test/MeshWeaver.Compiler.Pipeline.Test.csproj -c Release --no-build --no-restore --filter 'FullyQualifiedName~CompileCacheInputFreshnessTest|FullyQualifiedName~CompileSourceReadIdentityTest|FullyQualifiedName~CacheHitStampsTheSameRecordTest|FullyQualifiedName~ContentKeyReevaluationTest|FullyQualifiedName~GeneratedInputIdentityTest' --logger trx
cd test/MeshWeaver.Compiler.Pipeline.Test/bin/Release/net10.0
dotnet MeshWeaver.Compiler.Pipeline.Test.dll -culture en-US -trx compiler-suite-en-US.trx -showLiveOutput -longRunning 60

The local sighting and what it does not prove

The local 8339 Store preview supplied a separate production-code observation: a new native test class was present in a PDB-matched emitted source, while its test-area registration remained old. A later forced Compile at 04:43:49 returned a disk-cache hit on the 04:41 DLL and failed while pinning an unloading assembly context. A separately authorized normal retry later loaded the same DLL; its PDB still showed the old registration while the live/current and compiled source stamps agreed. The cache/compiler files were byte-identical at that preview's core 508aeb6 and the regression baseline 6231c4da.

The executed regression proves the cache's stale-byte/current-stamp mismatch. It does not establish which import/query boundary first supplied the preview's mixed snapshot, nor which operation caused the three failed scan-pin attempts. The unloading failure is a separate lifecycle investigation. Recycle unloads contexts but deliberately retains disk artifacts, so its success alone cannot certify source freshness.

Original local evidence is retained at /private/tmp/roadmap-compile-cache-investigation (before.trx, before-test.log, before-receipt.json, exact baseline source copies and subsequent test receipts). The preview's PDB and activity receipts remain under /private/tmp/coupon-input-preview-adoption-plan/post-cta-retry-pdb. These paths are local evidence, not dependencies of the committed test. No deployment or production verification is claimed here.