Full SDK header annotation migration (SDK28000 WIP)

.gitattributes

@@ -1,2 +1,3 @@
 RecompiledIdlHeaders.zip filter=lfs diff=lfs merge=lfs -text
 Windows.Win32.winmd filter=lfs diff=lfs merge=lfs -text
+generation/WinSDK/patches/**/*.patch -text

.github/copilot-instructions.md

@@ -74,7 +74,7 @@ Each arch job scrapes ALL partitions for that architecture.
 
 **Import discovery before header overrides:** Before adding `_Win32_ImportLibrary_` or a manual DLL mapping, inspect the selected SDK's import libraries and the library-processing set in `scripts\CreateProcLibMappingForAllLibs.ps1`. Use `WinmdUtils dumpImports --lib <path>` to verify the actual symbols and DLL, not just the library filename. If a matching `.lib` was omitted or its imports were missed, fix shared discovery instead of adding a redundant header annotation. Use an authored override only when no authoritative SDK library mapping is available or the native contract genuinely requires an override; record the SDK/library search evidence.
 
-**Header-annotation migration:** Follow `docs\copilot\plans\header-annotation-migration-playbook.md`. Classify the original artifact before editing: patch authored `.h` files directly, but patch original `.idl` and regenerate all affected outputs for IDL-owned declarations. Never patch a generated `.h` or independently patch the same shared artifact in different partitions. Use exact qualified tooling/pilot pins and keep raw OLD-to-NEW differences visible.
+**Header-annotation migration:** Follow `docs\copilot\plans\header-annotation-migration-playbook.md`. Annotate the definition consumed by the actual SDK preparation. If preparation runs MIDL to regenerate that header, patch the owning IDL input before MIDL and regenerate the affected outputs. Otherwise, patch the consumed shipped SDK header in the established patch phase, even if it has a generated banner or corresponding IDL/ODL is present but not regenerated. An external IDL/ODL import or version match is not a prerequisite for that header patch. Do not edit the SDK mirror or prepared outputs directly, or independently patch the same shared artifact in different partitions. Use exact qualified tooling/pilot pins and keep raw OLD-to-NEW differences visible.
 
 **Winmd equivalence:** After any change, the winmd must be compared against the baseline. The `NoSuggestedRemappings` test in `Windows.Win32.Tests` validates that no new unhandled remappings were introduced. API diffs are automatically posted as PR comments by CI. Run `.\scripts\DiffWinmdToBaseline.ps1` locally to see a full C# declaration diff against the last release.
 

.github/tests/pr-api-diff-comment.test.cjs

@@ -180,6 +180,11 @@ test('both artifact uploads still include the generated patch', () => {
   }
 });
 
+test('full SDK scrapes have a bounded two-hour budget while build and test keep one hour', () => {
+  assert.match(validationJob('scrape'), /^    timeout-minutes: 120$/m);
+  assert.match(validationJob('build-test'), /^    timeout-minutes: 60$/m);
+});
+
 for (const [jobName, uploadCount] of [['scrape', 1], ['build-test', 2]]) {
   test(`${jobName} always-run local uploads require a successful checkout`, () => {
     const job = validationJob(jobName);

.github/workflows/pr-validation.yml

@@ -98,7 +98,7 @@ jobs:
     needs: [changes]
     if: needs.changes.outputs.build_needed == 'true'
     runs-on: windows-2022
-    timeout-minutes: 60
+    timeout-minutes: 120
 
     strategy:
       fail-fast: false

docs/copilot/header-annotation-decisions.md

@@ -1,13 +1,59 @@
 # Header annotation migration: problems and decisions
 
-Updated: 2026-10-23 (UTC).
+Updated: 2026-10-25 (UTC).
 
 This is the running record of migration blockers, their evidence, and the decisions
 made while resolving them. The goal is to express win32metadata's intended API
 metadata faithfully in the selected SDK's original headers and IDL. A zero-difference
 OLD-to-NEW WinMD is not a goal. Breaking changes and unexplained differences remain
 visible for review.
 
+## Constant conflict policy
+
+Use `E_NOTFOUND = 0x80070490` from `um\devicetopology.h` as canonical.
+Resolve competing loose constant projections by header priority and first-wins
+selection. Give that header explicit priority; preserve the actual project
+partition, traversed-header and native declaration order for unlisted headers.
+Keep source-bound group ownership distinct from choosing between unrelated
+native definitions.
+
+Produce warnings and a complete conflict dump for final review. Retain every
+candidate's native identity, header and source location, architecture, type,
+storage and raw value, with the winner and selection reason. Include discarded
+definitions even when their values match. The competing definitions remain
+distinct native contracts, and raw OLD-to-NEW differences remain visible.
+
+## Full generation and green PR gates
+
+Produce the complete ordinary SDK image and its unfiltered same-SDK OLD-to-NEW
+comparison using one coherent published toolchain and prepared-header identity.
+Do not wait for individual native or physical investigations, projected-storage
+changes, or character-projection fixes before running the full scrape, constant
+extraction, manual-source composition, and emission.
+
+Both stacked PRs must pass their exact-head normal CI gates. Parser, compiler,
+emitter, integrity, and regression-test failures remain blocking and require
+repairs. Do not disable jobs, blanket-skip tests, suppress diagnostics, or report
+an incomplete image as successful.
+
+Keep the migration-difference report separate from those build gates. Known
+`PSTR` versus `CHAR*` projection and `VARENUM` storage differences are nonblocking
+follow-up findings for image production; they are not equality or qualified
+compatibility. Preserve the full raw differences and independent Const, import,
+flag, and parameter-sequence evidence. Retain native, SAL, and coverage findings
+without turning each investigation into a prerequisite for the next full image.
+After establishing green normal CI, make further repairs with focused regression
+coverage and a fresh normal CI run.
+
+For original IDL unavailable in the shipped SDK, ordinary tests enforce exact
+deferred-fact accounting rather than unconditional migration completion. The
+report-only `tests\SdkAnnotationCoverage.Tests\SourceBlockedMetadata.json` ledger
+retains all 30 historical units: 26 availability values, three associations and
+`IdentityUpdateEvent` Flags. It supplies no emitter input or semantic fallback.
+Missing entries, altered values, generated-header patches and false coverage or
+preservation claims remain failures. Current physical projection is unqualified;
+raw losses remain visible in the bulk comparison.
+
 ## Completion plan and next review
 
 The next run follows a reviewed repair of the header-only source migration.
@@ -66,8 +112,10 @@ unchanged preparation test fails with MIDL9008 beneath the long session-artifact
 prefix and passes beneath `artifacts\t-002`; no preparation success policy or
 fixture spaces/quoting is changed. Source-state tests inspect effective global
 plus partition annotation settings and source-owned facts, not retired sidecar
-rows. Retiring a row does not prove source coverage: missing original-IDL floors
-and native associations remain explicit test failures, not accepted baselines.
+rows. Retiring a row does not prove source coverage. Missing original-IDL floors,
+native associations and Flags remain exactly inventoried, deferred and
+unqualified; ordinary accounting regressions reject unaccounted changes or false
+coverage claims without making incomplete migration a generation prerequisite.
 
 The manual-type inventory exposed replacement-era exclusions that survived
 deletion of their handwritten providers. The frozen source packet retires ten