Round 2 (9ba8f013e) passed the retail and production lenses and every gate
and mutation was reproduced by all three; the gate-honesty lens failed it
on comment truth only: the terrace-edge Fact repeated round 1's false
'PC already matched' sentence, the sixth Fact claimed a 12/4 guard split
and a preamble seed the gate does not evaluate, and WalkVisibilityMath's
scan note cited a pin that does not exist and a callers argument that
was not the real justification. A third implementer round is barred by
the campaign rule, so the lead corrected the sentences at landing.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Round 1 (2bd353234) was three-lens reviewed: retail-faithful PASS,
production PASS, gate-honesty FAIL. Every finding landed on the
terrace-edge KnownFailure row plus one pin gap. This is the last round.
R2-1 - gate honesty. WalkTraceConformanceTests.AlphaDepthTranscript.cs's
own KnownFailure doc comment claimed "the PC sequence matches exactly";
it never did - Assert.Equal(expected.PmEvents, actualPm) threw first, so
the PC Assert.Equal line never ran. RunAlphaDepthTranscriptGate now
computes BOTH diffs (SequenceEqual, not Assert.Equal) before asserting
either, and Assert.Fail's message prints both sequences together
whenever either diverges - a written claim about one sequence is now
only ever one this method actually evaluated. Re-running the OLD
(unseeded, cold PortalsDrawnCount=0) code path through this new combined
assertion shows terrace-edge's real divergence was WORSE than round 1
believed: both PM and PC diverged (PC expected counter=2, actual
counter=0) - round 1's "PC matches" claim was not just untested, it was
false.
R2-2 - the harness fix. Retail's portalsDrawnCount (wo(008719b4)) is a
PERSISTENT session global; each capture's own FIRST PM or PC line
(cathedral-arrival 0, cathedral-leak 0, foundry-deep 1,
holtburg-doorway-still 2, terrace-edge 2, cathedral-stair-arch 8) already
carries the pre-capture value - terrace-edge's own sample sits in the
"F 1" preamble text a cdb session prints before the first parsed frame
marker (line 1209, ahead of "F 1" at line 1298), so no amount of
replaying the CAPTURED frames from a cold start can ever reach it (this
fixture's own turns never touch the counter - every PC line reads
ov=0). New WalkAlphaDepthTrace.LoadInitialCounter(root, fixtureName)
scans a fixture's raw lines, ignoring frame boundaries, for the first
PM/PC match and returns its counter. RunAlphaDepthTranscriptGate seeds
WalkFrameDriver.PortalsDrawnCount (already an internal, test-assembly-
settable field) from this value, runs ONE Collect+Replay pass, and
deletes the two-pass "priming" replay outright. Result: terrace-edge
turns GREEN with no KnownFailure tag; every other pose stays exact
(a fixture with real steady-state seal activity computes the identical
value a priming pass used to converge to, since a steady state is by
definition unchanged by one more pass). No pose diverges after seeding,
so the packet's "write the divergence into section 9" fallback never
triggers - nothing needed there this round.
R2-3 - the SOME-vs-SAME-plane discriminator. T2's existing five cases
never separated retail's real quantifier ("every vertex lies on the SAME
one of the four +/-12 planes") from a weaker "every vertex lies on SOME
plane" (per-vertex OR across the four planes, ANDed across vertices) -
the existing split-plane case's third vertex sits on NO plane at all, so
both readings admit it for the same reason. New case in
WalkVisibilityMathTests.cs
(Boundary_guard_admits_a_polygon_whose_every_vertex_is_on_SOME_plane_but_not_the_SAME_one)
and a fourth sink.OnPunchGeometry call in WalkFrameDriverTests.cs's
OnPunchGeometry_RejectsOnlyWhenEveryVertexSharesOnePlane_ButPunchesAnyOtherShape
pin close the gap: [(12,0,3),(0,12,3),(12,5,3)] - every vertex touches
SOME plane (vertex1/3 on x=+12, vertex2 on y=+12) but no SINGLE plane
holds all three (x=+12 cleared by vertex2's x=0, y=+12 cleared by
vertex1's y=0) - retail's real four-accumulator guard admits it; the
weaker per-vertex-OR form rejects it.
R2-4 - the sixth pose. Cherry-picked 4f12cd06b (walk+alphadepth capture)
and 3a6aa7eef (pmverts capture) from claude/campaign-w-retail-frame-walk
onto 2bd353234 - both applied clean, no conflicts (README.md merged
without a manual row). New
AlphaDepthTranscript_CathedralStairArch_MatchesRetailFrame2 Fact: root
0xF4180114, seven outside views, sixteen mode=1 building punches (all
counterBefore=0008 - far-Z punches never touch the counter) of which
twelve are guard-rejected on local y=12.000, then eight mode=0 exit
seals with counterBefore running 0..7 against the root's own ov=7
read-then-zero, settling this fixture's own steady-state counter at 8 -
exactly the value LoadInitialCounter reads back out of its own preamble.
PASS.
R2-5 - comment truth. (a) RetailPViewPassExecutor.cs:
DrawPortalDepthWrite's and DrawExitPortalMask's summaries corrected -
neither the retired ">=3 vertices" pre-filter nor "always matches what
reached the GPU" describe F2's actual count-before-clip behavior any
more. (b) The F2 pin's doc comment
(RetailPViewPassExecutorTests.DrawExitPortalMask_CountsAnUnclippableTwoVertexPolygon_ButDrawsNothing)
named TWO alternative mutations as if both failed; verified only one
does (see mutations below) - the doc comment now says so explicitly and
keeps only the pre-filter mutation. (c) WalkVisibilityMath.cs's doc
comment stated DAT-scan figures as bare facts; independently re-verified
this round with a freshly-written scan (own code, walking the same
production EnvCell.CellPortals/CellStruct.Polygons shape
WalkCellFactory itself uses, over the full 0x0000-0xFFFF landblock-prefix
range) - every cited figure reproduced bit-exact (3,405 landblocks with
cells; 1,854,237 portal polygons; 2,889 any-vertex; 2,163 all-on-plane,
all exit, 0 interior; 16,939 exit portals total, 1,837,298 interior) -
so the doc comment now states the counting rule (id range, what's
counted, how unresolvable polygons/cells are skipped) alongside every
figure instead of leaving them unqualified; landblocksWithCells=3,405
also independently matches CellStructSurfaceConstructionInstalledDatTests'
own OH2 walk of the identical range (that test's own golden-count
comment). The scratch scan itself is not part of this diff (deleted
after use, same as F1's own datScan). (d) This commit body does not
cite "register tests" as evidence - no test reads the register.
Register truth is the human read: no row changes, AD-119 (the C2
blend-vs-write-mask row) unchanged.
Mutations run this round (all verified against production code, then
reverted):
- R2-1/R2-2 (disable seeding, driver.PortalsDrawnCount = 0 instead of
initialCounter): AlphaDepthTranscript_TerraceEdge_MatchesRetailFrame2
fails - "terrace-edge.alphadepth: PM DIVERGES - expected
[(mode=1,counterBefore=2), (mode=1,counterBefore=2)], actual
[(mode=1,counterBefore=0), (mode=1,counterBefore=0)]; PC DIVERGES -
expected [(ov=0,counter=2,fc=0), (ov=0,counter=2,fc=0)], actual
[(ov=0,counter=0,fc=0), (ov=0,counter=0,fc=0)]" - proving both the
seeding fix (R2-2) and the honest dual-sequence message (R2-1) at
once.
- R2-3 (replace the four per-plane accumulators with a single
per-vertex "on ANY plane" test ANDed across vertices):
Boundary_guard_admits_a_polygon_whose_every_vertex_is_on_SOME_plane_but_not_the_SAME_one
fails, Assert.False() Expected:False Actual:True;
OnPunchGeometry_RejectsOnlyWhenEveryVertexSharesOnePlane_ButPunchesAnyOtherShape
fails, Assert.Equal() Expected:3 Actual:2.
- R2-5b, mutation ONE (restore the old `localVertices.Length < 3`
pre-filter ahead of the guard):
DrawExitPortalMask_CountsAnUnclippableTwoVertexPolygon_ButDrawsNothing
fails, Assert.Equal() Expected:1 Actual:0 - kept as the pin's cited
mutation.
- R2-5b, mutation TWO (move `submitted++` to AFTER
`_portalDepthMask.DrawDepthFan(...)` instead of before): same test
STILL PASSES - DrawDepthFan has no effect on the local `submitted`
counter, so the final returned value is identical regardless of
which side of that call the increment sits on. Confirmed unobservable
exactly as R2-5b required; the doc comment no longer names this as a
failing mutation.
Gates: dotnet build tests/AcDream.App.Tests -c Release -> 0 warnings/0
errors. dotnet build src/AcDream.App -c Release -> 0 warnings/0 errors.
Hermetic lane (Lane!=InstalledDat&Lane!=PreparedPackage&Lane!=Live&
Lane!=Manual&Lane!=Timing&Lane!=Windows&Lane!=Linux&Lane!=SystemFont&
Purpose!=Diagnostic&Status!=KnownFailure): 6830/6830 passed (+1 over
round 1's 6829, the new T2 case). InstalledDat lane
(ACDREAM_RUN_INSTALLED_DAT_TESTS=1, --filter Lane=InstalledDat): 254
total (+1 over round 1's 253, the new sixth Fact), 249 passed, 4 failed -
exactly the pre-existing four (LayoutImporterMediaBearingChildSweepTests
and LayoutImporterInvisibleSweepTests, both #383;
TowerAscentReplayTests' KnownFailure; Oh_doorway_still_first_frame_diff,
#458) plus the pre-existing WalkLookInGateSweepTests skip - terrace-edge
no longer fails, no new divergence anywhere else. All six
AlphaDepthTranscript_*_MatchesRetailFrame2 Facts: 6/6 passed standalone.
Shader classes (VulkanShaderDescriptorContractTests +
VulkanShaderManifestTests + RenderPackSpirvValidatorTests +
SkyVertexLayoutTests): 35/35. RetailPViewPassExecutorTests +
WalkFrameDriverTests (full classes, regression check on the R2-3/R2-5b
edits): 57/57.
Commits: two cherry-picks (4f12cd06b, 3a6aa7eef) from
claude/campaign-w-retail-frame-walk, both clean, plus this one fix
commit on top of 2bd353234.
Register: no new AD row; AD-119 unchanged (human read, not a test -
see R2-5d above).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Three-lens review of c7ab5b6d8 found the round-0 port of
DrawPortalPolyInternal's degenerate-input guard inverted retail's
quantifier: it rejected a polygon when ANY vertex sat on ANY of the
four +/-12 planes. The PDB-paired bytes (VA 0x59BCD6-0x59BD66: four
per-plane fld/fcomp 12.0/fnstsw/test ah,0x44/jnp sequences, each
accumulating a "every vertex on this plane" boolean, cleared by the
first vertex NOT on it) and the named decomp's own nested-if reading
of the post-loop branch (0x59bd42-0x59bd6c) agree: retail rejects only
when EVERY local vertex lies on the SAME one of the four planes — a
polygon degenerate onto a plane, not one that merely touches a plane
at one vertex.
F1 — the quantifier. WalkVisibilityMath.IsRejectedByPortalPolygonBoundaryGuard
now tracks four booleans (everyVertexOnPlusX/MinusX/PlusY/MinusY), each
starting true and cleared by the first vertex that fails its own plane
test; the guard rejects iff any of the four survives the whole vertex
loop. T2 (WalkVisibilityMathTests.cs) is rewritten to the five §7
cases: every vertex on x=+12 -> rejected; every vertex on y=-12 ->
rejected; one vertex on x=+12, rest inside -> KEPT; every vertex at
11.999 -> kept; vertices split across x=+12 and y=+12 (no common
plane) -> kept. WalkFrameDriverTests' punch-side pin
(OnPunchGeometry_RejectsOnlyWhenEveryVertexSharesOnePlane_ButPunchesAnyOtherShape)
carries the same correction — its old "one vertex on the plane"
polygon was wrongly rejected under round 0; it punches now, and a new
all-on-plane polygon is added as the actual rejected case. The doc
comment's "essentially never" sentence is deleted and replaced with a
fresh DAT-scan citation (see datScan below).
MUTATION (verified): restore the any-vertex form -> the one-vertex and
split-plane admit cases fail (Assert.False Expected:False Actual:True,
x2; the WalkFrameDriverTests punch pin's Assert.Equal Expected:2
Actual:1). Delete the guard (always return false) -> the all-on-plane
reject cases fail (Assert.True Expected:True Actual:False, x3 —
plusX-plane, minusY-plane, and the vacuous-empty-polygon pin; the
punch pin's Assert.Equal Expected:2 Actual:3).
datScan (DAT counts backing F1's doc comment): a throwaway xunit test
(deleted after this commit — not part of the diff) walked the full
0x0000-0xFFFF landblock-prefix space (LandBlockInfo/EnvCell carry no
usable DatCollection.GetAllIdsOfType range — both are landblock-
relative — matching the method CellStructSurfaceConstructionInstalledDatTests
already established), resolved each EnvCell's Environment/CellStruct,
and for every CellPortal's polygon tested each vertex's local x/y
against +/-12. Against the installed client_cell_1.dat: 3,405
landblocks with cells, 1,854,237 total portal polygons. 2,889 polygons
have at least one vertex on a +/-12 plane; 2,163 have EVERY vertex on
the SAME plane — all 2,163 are EXIT portals (OtherCellId==0xFFFF, out
of 16,939 exit portals total); 0 interior (non-exit) portals ever
qualify (interior: 1,837,298 total, 340 any-vertex, 0 all-on-plane).
Duration ~15s. This exactly reproduces the review's cited numbers.
F2 — seal-path count order. RetailPViewPassExecutor's exit-seal loop
dropped the `localVertices.Length < 3` pre-filter that ran BEFORE both
the boundary guard and the `submitted` increment (round 0's order:
length check -> guard -> transform -> draw -> count). Retail's real
order is guard -> transform -> clip -> count, with the counter
incremented BEFORE polyClipFinish runs (0x59BD70-0x59BD74 precedes
0x59BDB0) — the counter records accepted ATTEMPTS, not successful GPU
fans. `submitted++` now runs immediately after the guard-gated
transform+EmitSeamMask, before DrawDepthFan (whose own `< 3` guard is
this port's stand-in for retail's post-clip `var_4 >= 3` check). A new
pin (RetailPViewPassExecutorTests.DrawExitPortalMask_CountsAnUnclippableTwoVertexPolygon_ButDrawsNothing)
constructs a real PortalDepthMaskRenderer (RecordingGpuDevice-backed)
and a constructor-bypassed executor (RuntimeHelpers.GetUninitializedObject,
the same pattern WalkOutsideViewReassemblyTests already uses) with a
2-vertex ordinary-coordinate portal polygon: DrawExitPortalMask returns
submitted=1 while zero draw calls reach the device.
MUTATION (verified): restore the `< 3` pre-filter ahead of the guard
-> the pin's Assert.Equal fails, Expected:1 Actual:0.
F3 — the depth-event transcript gate, automated. New
WalkAlphaDepthTrace.cs (beside WalkOracleTrace.cs) parses the four
`PM poly=<ptr> mode=<0|1> counterBefore=<hex>` / `PC ov=<n>
counter=<hex> fc=<0|1>` lines per oh-capture/*.alphadepth.log frame,
using the same F-marker frame delimiting as WalkOracleTrace. New
WalkTraceConformanceTests.AlphaDepthTranscript.cs (beside
WalkTraceConformanceTests.cs, which is now `partial`) replays each
pose's frame 2 through the real WalkFrameDriver/RetailFrameWalk over
WalkLandscapeDatBuilder-built installed-DAT world data, with:
- AlphaDepthCollectSink: an IWalkEventSink decorator forwarding every
hook to the real driver unchanged, while recording mode=1 punch
events (OnPunchGeometry, counterBefore = driver.PortalsDrawnCount
at COLLECT time — Collect finishes entirely before Replay starts,
so reading the counter at Replay time would already see the SAME
turn's post-reset value; holtburg-doorway-still's own capture
proves this: its punches show counterBefore=2 immediately followed
by its seals' counterBefore=0/1 in the identical turn) and every
PC tuple (Emit, WalkEventKind.DrawCells — fires for the root's own
turn AND every building look-in's re-entrant call, matching
retail's breakpoint not distinguishing them).
- AlphaDepthLeaf: an IWalkFrameLeafRenderer whose DrawExitSeals
reproduces DrawPortalDepthWrite's real per-portal loop (F1's guard,
F2's count-before-clip order) directly over
driver.InteriorFloodCells / WalkCell.Portals / .PortalPolygons —
the identical data the render-side production path reads — since
the interface gives DrawExitSeals no cell/portal parameter at all.
- Two Collect+Replay passes per pose (a fresh driver reaches the SAME
steady-state a running session already carries by the time
retail's capture began — holtburg-doorway-still and foundry-deep
both show a stable per-frame accepted-seal count that regenerates
identically every captured frame): the first pass's own output is
discarded, only the second is compared.
Five Facts (one per pose — the spec's "four" undercounts the five
*.alphadepth.log files actually present in oh-capture/):
AlphaDepthTranscript_CathedralArrival_MatchesRetailFrame2 — PASS
AlphaDepthTranscript_CathedralLeak_MatchesRetailFrame2 — PASS
AlphaDepthTranscript_FoundryDeep_MatchesRetailFrame2 — PASS
AlphaDepthTranscript_HoltburgDoorwayStill_MatchesRetailFrame2 — PASS
(proves the incrementing 0/1/2 accepted-seal sequence AND the
pre-reset counterBefore=2 punches, end to end, against real DAT
portal data)
AlphaDepthTranscript_TerraceEdge_MatchesRetailFrame2 — KnownFailure
(tagged [Trait("Status","KnownFailure")]): terrace-edge never runs
a qualifying interior turn in its own capture (every PC line reads
ov=0) — its punches' counterBefore=2 is a session value carried
over from BEFORE the capture even started (the file's own pre-"F 1"
content already reads 2; nothing inside the captured frames ever
touches the counter). A fresh two-pass replay from
PortalsDrawnCount=0 has no mechanism to derive that leftover value.
Printed divergence: expected [(mode=1,counterBefore=2) x2], actual
[(mode=1,counterBefore=0) x2]; the PC sequence (ov=0 both times)
matches exactly. An initial-condition gap in the fixture, not a
guard/count defect — the identical harness reproduces holtburg-
doorway-still's and foundry-deep's own steady states exactly from
cold starts.
MUTATION (verified): swap AlphaDepthLeaf's recorded seal mode from 0
to 1 -> AlphaDepthTranscript_HoltburgDoorwayStill_MatchesRetailFrame2
fails (Assert.Equal Collections differ at pos 5: expected
(0,0),(0,1)... actual (1,0),(1,1)...).
DispatcherFixture and DrawScope (WalkFrameDriverTests.cs) are widened
from private to internal so the new gate can drive a real
WbDrawDispatcher without duplicating that fixture.
F4 — manifest comment truth. VulkanShaderManifestTests.cs's
portal_depth.vert.spv re-pin comment had the ULP direction backward
(claimed the OLD 0x3F7FFFFE literal was NEARER the camera than
retail's real 0x3F7FFFEF constant; portal_depth.vert's own shader-
source comment already had this correct — FARTHER, larger z/w) and
cited T1 as living in WalkVisibilityMathTests when
PortalDepthVert_FarPunchConstant_MatchesRetailExactBits is in this
same file. Both corrected.
Gates: dotnet build tests/AcDream.App.Tests -c Release -> 0 warnings/0
errors. Hermetic lane (Lane!=InstalledDat&...&Status!=KnownFailure):
6829/6829 passed. InstalledDat lane (real client_cell_1.dat): 253
total, 247 passed, 5 failed — exactly the two #383 layout tests
(LayoutImporterMediaBearingChildSweepTests/LayoutImporterInvisibleSweepTests),
TowerAscentReplayTests' KnownFailure, Oh_doorway_still_first_frame_diff
(#458), and this round's own AlphaDepthTranscript_TerraceEdge
KnownFailure — no other divergence. Shader tests (VulkanShaderDescriptorContractTests
+ VulkanShaderManifestTests + RenderPackSpirvValidatorTests +
SkyVertexLayoutTests): 35/35. Register/divergence tests: 52/52.
dotnet build src/AcDream.App -c Release: 0 warnings/0 errors.
Register: no new AD row. F1/F2 are straight bug fixes toward retail
fidelity (not deviations); AD-119 (the C2 blend-vs-write-mask row)
does not describe the count order, so its neighbour needed no
correction per F2's own instruction.
Deviation from the fix-round spec's literal text: §7 F1 said "swap the
any-vertex hits instead of the four per-plane accumulators" as the
split-plane case's mutation — the code actually ported uses per-vertex
`if (x != 12f) flag = false;` clears, not a per-vertex OR; the mutation
verified above (restore the any-vertex FORM entirely) exercises the
same defect class and is the mutation actually run. §7 also names
"four" alphadepth.log fixtures where five exist on disk
(cathedral-arrival, cathedral-leak, foundry-deep,
holtburg-doorway-still, terrace-edge); all five are gated rather than
four, with cathedral-leak an unremarked addition alongside the four
the spec names.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Erik