docs(overhaul): contract s5 closeout witness harness
This commit is contained in:
Erik
parent
aa51cb88b0
commit
4237a6b2b3
2 changed files with 182 additions and 1 deletions
|
|
@ -3312,3 +3312,184 @@ production/gate-honesty lens review filters, provenance, manifests, negative
|
|||
results, and absence of retries. Any missing witness becomes a separately
|
||||
contracted narrow harness chunk; otherwise C1 closes and C2 runs the connected
|
||||
lifecycle/performance program.
|
||||
|
||||
## 37. S5 closeout C1 result and C1a narrow harness contract — 2026-09-05
|
||||
|
||||
### 37.1 C1 result: the offline matrix is green and two evidence gaps are real
|
||||
|
||||
Section 36's code base was `29c05e48c8379678603efb8715f86ede04c775ea`;
|
||||
the executed clean contract tip was
|
||||
`aa51cb88b00c843f87aa48047f9f3ce939755909`. One OpenAI implementer ran each
|
||||
official gate exactly once, with no retry, edit, client launch, package bake,
|
||||
surrogate witness, or cleanup:
|
||||
|
||||
```powershell
|
||||
& tools/run-release-gate.ps1 `
|
||||
-ArtifactsDirectory artifacts/s5-closeout-c1-hermetic-aa51cb88b `
|
||||
-TestFilter 'Lane!=InstalledDat&Lane!=Linux&Lane!=Manual&Lane!=Timing&Lane!=Live&Lane!=PreparedPackage'
|
||||
|
||||
$env:ACDREAM_RUN_INSTALLED_DAT_TESTS='1'
|
||||
$env:ACDREAM_DAT_DIR='C:\Users\erikn\Documents\Asheron''s Call'
|
||||
& tools/run-release-gate.ps1 `
|
||||
-ArtifactsDirectory artifacts/s5-closeout-c1-installeddat-aa51cb88b `
|
||||
-SkipRestore -SkipBuild `
|
||||
-TestFilter 'Lane=InstalledDat&Purpose!=Diagnostic'
|
||||
```
|
||||
|
||||
The two temporary environment variables were removed after the second gate.
|
||||
|
||||
- literal hermetic artifact
|
||||
`artifacts/s5-closeout-c1-hermetic-aa51cb88b`: Release build 0 warnings / 0
|
||||
errors, 16,942 passed / 0 failed / 0 skipped, manifest 32/32;
|
||||
- canonical InstalledDat artifact
|
||||
`artifacts/s5-closeout-c1-installeddat-aa51cb88b`: 368 passed / the accepted
|
||||
nine failures / one accepted look-in skip, manifest 30/30.
|
||||
|
||||
The nonpassing identities match S5-c5 exactly. The skip is
|
||||
`WalkLookInGateSweepTests.Sweep_the_lookin_gate_decodes_against_the_street_fixture`.
|
||||
The failures are
|
||||
`WalkTraceConformanceTests.AlphaFlushCounts_CathedralArrival_MatchesRetailFrame2`,
|
||||
`WalkTraceConformanceTests.AlphaFlushCounts_CathedralLeak_MatchesRetailFrame2`,
|
||||
`WalkTraceConformanceTests.AlphaFlushCounts_CathedralStairArch_MatchesRetailFrame2`,
|
||||
`WalkTraceConformanceTests.AlphaFlushCounts_FoundryDeep_MatchesRetailFrame2`,
|
||||
`WalkTraceConformanceTests.AlphaFlushCounts_HoltburgDoorwayStill_MatchesRetailFrame2`,
|
||||
`WalkTraceConformanceTests.AlphaFlushCounts_TerraceEdge_MatchesRetailFrame2`,
|
||||
`WalkTraceConformanceTests.Oh_doorway_still_first_frame_diff`,
|
||||
`LayoutImporterInvisibleSweepTests.EveryAuthoredInvisibleWidget_StartsHiddenAcrossAllLayouts`,
|
||||
and
|
||||
`LayoutImporterMediaBearingChildSweepTests.MainGameUiAndChatInput_MediaBearingChildrenNowBuildAsRealWidgets`.
|
||||
The two summary JSON files identify the exact HEAD/branch, report
|
||||
`WorktreeDirty=false`, and their SHA manifests independently re-verify.
|
||||
|
||||
The TRX map covers the focused App, Content, Core, Runtime, and Launcher
|
||||
requirements, including `BeginWalkFrame`, outside-view reassembly, retail
|
||||
membership, FIFO/barrier/order, material, degrade, visibility consumers,
|
||||
reset/abort/dispose, architecture guards, and the real warmed-allocation pins.
|
||||
It also proved two missing witnesses rather than hiding them:
|
||||
|
||||
1. `cathedral-stair-arch.walk.log` is the only committed OH walk transcript
|
||||
with no test consumer. Its `.alphadepth` sibling is consumed, but that is a
|
||||
different product and cannot stand in for the walk transcript.
|
||||
2. No executable gate hashes a canonical runtime geometry product together
|
||||
with the committed retail membership product. The existing installed-DAT
|
||||
CellStruct hash covers counts/metadata rather than vertex/index values,
|
||||
transforms, or membership; the scene digest is synthetic; and the existing
|
||||
retail comparator uses test-assembled terrain and copied publication
|
||||
dispatch. Equal source, package, log, or TRX hashes are not substitutes.
|
||||
|
||||
These are test-harness gaps, not evidence of a renderer behavior defect. C1
|
||||
therefore closes as an honest offline inventory and C1a below is the required
|
||||
narrow correction before connected/performance C2.
|
||||
|
||||
### 37.2 C1a scope: consume the sixth walk fixture
|
||||
|
||||
Edit only the following walk tests for this half:
|
||||
|
||||
- add `(OhRoot, "cathedral-stair-arch.walk")` to
|
||||
`WalkOracleTraceTests.AllFixtures`, so the ordinary complete-frame parser
|
||||
contract consumes the checked-in walk log; and
|
||||
- add a dedicated installed-DAT conformance test shaped like
|
||||
`Oh_foundry_deep_reproduces_every_complete_frame_exactly` in
|
||||
`WalkTraceConformanceTests`. Replay every complete marked stair-arch frame
|
||||
and compare the complete eight-kind `Signature8` exactly. A row in
|
||||
`Still_fixture_first_frame_reproduces_exactly` is insufficient because that
|
||||
theory checks only `frames[1]`. Do not compare only event counts or borrow
|
||||
the alpha-depth transcript.
|
||||
|
||||
The new row must load the existing committed file. Do not recapture, rewrite,
|
||||
normalize, or manufacture a fixture. If the production replay cannot reproduce
|
||||
the row under the same adapter used by the other OH rows, return the exact
|
||||
first divergence as a finding; do not weaken the signature.
|
||||
|
||||
### 37.3 C1a scope: canonical runtime geometry/membership product witness
|
||||
|
||||
Add one InstalledDat App test class, plus at most one test-only data adapter if
|
||||
the class cannot remain readable. Production source, fixtures, package files,
|
||||
gate scripts, shaders, and register documents are out of scope. Each of two
|
||||
independent runs inside the test must open a fresh installed `DatCollection`
|
||||
and construct fresh extraction, physics, and membership owners. Use canonical
|
||||
cathedral landblock `0xF418FFFF` with origin `(0xF4,0x18)`, real region height
|
||||
data, complete EnvCells, and independent interior statics.
|
||||
|
||||
Build through the real `LandblockBuildFactory.Build`. A test-only
|
||||
`IPreparedCollisionSource` may translate installed GfxObj, Setup, EnvCell, and
|
||||
CellStruct records through the existing `FlatCollisionAssetBuilder`
|
||||
flattening methods, but it may not compute membership. Read actual ordered
|
||||
EnvCell shell geometry through `DatPreparedAssetSource.Read` with the same
|
||||
`PreparedAssetRequest.EnvCellGeometry` product consumed by
|
||||
`ObjectMeshManager`; read every referenced Setup/GfxObj product through the
|
||||
same production extractor, including ordered Setup part placements. Publish
|
||||
membership through the real no-window production sequence on
|
||||
`LandblockPhysicsContentBuilder`: `BuildTerrainSurface`, `PublishPreparedCells`,
|
||||
`CacheBuildings`, `CachePreparedObjects`, `PhysicsEngine.AddLandblock`, then
|
||||
`PublishStaticCollision`. The test must not copy this branch dispatch or call
|
||||
`ShadowObjectRegistry.Register*` to construct the expected membership.
|
||||
|
||||
Serialize a version-tagged binary product with explicit lengths and raw IEEE
|
||||
float bits, then SHA-256 it. The product must include:
|
||||
|
||||
- geometry identities; every emitted vertex position, normal, and UV; every
|
||||
index; semantic subset order and `SourceSurfaceIndex`, `RawSurfaceType`,
|
||||
`RetailSurfaceMask`, and `IsCellShell`; ordered Setup part ids and full
|
||||
transforms; and the actual placed cell/static transforms;
|
||||
- for every published owner, its `RetailCellArrayRoute`, its ordered CELLARRAY
|
||||
from `TryGetRetailCellArray`, and every ordered per-cell `RetailPartEntry`
|
||||
from `GetRetailPartEntriesInCell`: `EntityId`, `PartIndex`, `GfxObjId`,
|
||||
`CellId`, and `ClipPlanesRequired`.
|
||||
|
||||
Canonicalize only outer maps whose production semantics are unordered.
|
||||
CELLARRAY order, Setup-part order, subset order, index order, publication
|
||||
order, and per-cell entry order remain exact. Assert real shell geometry for
|
||||
cell `0xF4180104` with nonzero vertices/subsets/indices. Assert the
|
||||
production-built cathedral ramp Setup `0x020009A2` exists and that its seven
|
||||
parts produce 21 entries across the exact ordered CELLARRAY
|
||||
`[0xF4180112, 0xF4180113, 0xF4180009]`, all with
|
||||
`ClipPlanesRequired=true`. Assert at least one multi-cell product and one
|
||||
`BoundingBox` route. Return the per-domain counts plus geometry, membership,
|
||||
and combined hashes.
|
||||
|
||||
Two fresh in-test runs must produce the same three hashes and exact counts.
|
||||
The focused test must also be run in two fresh test processes at the same
|
||||
clean commit, and its labeled digest/count output must agree. Equality only
|
||||
proves determinism, so pin the clean installed-DAT combined hash and the exact
|
||||
cathedral semantic facts above as correctness witnesses. If installed DAT
|
||||
identity differs from the campaign's accepted corpus, fail explicitly rather
|
||||
than refreshing the golden silently.
|
||||
|
||||
The retail meaning remains the already-landed S2 contract: whole-part-array
|
||||
CELLARRAY calculation and ordered part insertion, anchored by named retail
|
||||
`CPhysicsObj::AddPartToShadowCells` (`0x00510e00`),
|
||||
`CPhysicsObj::add_shadows_to_cells` (`0x00514ae0`),
|
||||
`CObjCell::find_cell_list` (`0x0052b4e0`), and `CELLARRAY::add_cell`
|
||||
(`0x006b4ff0`). C1a changes no AC algorithm.
|
||||
|
||||
### 37.4 Required discrimination, gates, and return
|
||||
|
||||
Before committing, run and restore these mutations one at a time, recording
|
||||
the exact first failing assertion:
|
||||
|
||||
1. change one serialized vertex float bit while preserving counts;
|
||||
2. swap two real subset indices while preserving counts;
|
||||
3. reverse two real ordered CELLARRAY cells or Setup parts before publication;
|
||||
4. flip the cathedral ramp's real `ClipPlanesRequired` value; and
|
||||
5. redirect the new stair-arch row to a nonexistent fixture.
|
||||
|
||||
Each product mutation must be reached by the canonical corpus and must fail
|
||||
the pinned product proof, not merely a self-comparison. Restore every file
|
||||
byte-for-byte and prove `git diff --check` clean. Return one test-only commit,
|
||||
the exact changed-file list, all mutation transcripts, a zero-warning Release
|
||||
solution build, the new fixture-consumer and product-hash focused tests, then
|
||||
fresh official literal hermetic and canonical InstalledDat artifacts with
|
||||
verified manifests and exact nonpassing identities. Any production edit,
|
||||
fixture edit, hash-field omission, copied membership construction, new
|
||||
nonpassing identity, or inability to form a non-vacuous real product stops the
|
||||
chunk and returns a finding. No graphical client is authorized for C1a.
|
||||
|
||||
Review is sequential. Retail/transcript pass 1 checks the exact committed
|
||||
fixture, `Signature8`, named-retail product/order meaning, field completeness,
|
||||
and golden discrimination. Production/gate-honesty pass 2 checks fresh-owner
|
||||
independence, real factory/mesh/publication reachability, absence of copied
|
||||
registration dispatch, canonicalization boundaries, InstalledDat lane
|
||||
provenance, manifest truth, and allowed scope. A failed lens gets one bounded
|
||||
fix contract; review stops only after pass 10. After C1a lands, C2 runs the
|
||||
connected lifecycle/performance closeout and G4 follows.
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue