Port retail's per-frame light collection: the point-light pool is built from ONLY the
currently-visible cells' lights, matching CObjCell::add_*_to_global_lights
(0x0052b350/0x0052b390) walked over CEnvCell::visible_cell_table (0x0052d410) — not a
flat world-space set capped at 128-nearest-camera.
- LightSource.CellId (retail insert_light arg6 -> RenderLight +0x6c); tagged at both
registration sites from entity.ParentCellId (live weenie fixtures + dat EnvCell statics).
- LightManager.BuildPointLightSnapshot(camPos, visibleCells): a light joins the pool iff
CellId==0 (viewer/global) or its cell is in the flood. 128 cap kept as a now-non-biting
backstop (retail's is 40 static + 7 dynamic, 0x0081ec94/8).
- Threaded via RetailPViewDrawContext.RebuildScopedLights, invoked in DrawInside after the
flood resolves prepareCells and before the draws (renderers select from the same
in-place-rebuilt PointSnapshot; EnvCellRenderer clears its per-cell cache each pass).
- [indoor-light] probe (ACDREAM_PROBE_INDOOR_LIGHT=1) dumps the scoped-pool SET COMPOSITION.
Un-skips LightManagerTests.PointSnapshot_HubScaleLightCount_ObjectSelectionIsCameraInvariant.
CORRECTION: the handoff called the camera-cap the "confirmed" #176/#177 mechanism. The probe
PROVES scoping works (291 Hub fixtures -> pool of 1-9, ~285 through-floor lights dropped/frame,
CellIds match the flood), but the user's VISUAL GATE showed BOTH symptoms unchanged. So pool
composition is NOT the cause. #176 real cause = an over-bright purple point light
(intensity=100, color 0.784,0,0.784 -- from [light-detail]); #177 = a portal-visibility miss
(stairs not drawn looking back). Both stay OPEN. This change is retail-faithful and retires the
camera-eviction latent bug; kept as such, not as the symptom fix. Register AP-85 corrected;
ISSUES #176/#177 re-diagnosed; render digest banner updated.
Decomp: insert_light 0x0054d1b0, minimize_object_lighting 0x0054d480, calc_point_light
0x0059c8b0; pseudocode docs/research/2026-07-06-a7-per-cell-lighting-pseudocode.md.
Suites green: Core 2595 + 2 skip, App 719 + 2 skip.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Dense-town before/after (frame profiler + user eyes): frame-time max
20-87ms -> 6-10ms, single-frame alloc spikes (30-75MB) eliminated,
gen2 GC 5-11/window -> ~0, user confirms the FPS counter now holds
steady instead of swinging. The residual steady ~1.6MB/frame (from the
deferred harder sites: EnvCell rebuild gate + physics Transition
pooling) is small gen0-only churn that no longer causes swings; those
are now an OPTIONAL follow-up that would also lower the median.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Completes the safe-batch (implementer finished this but died before
committing; coordinator independently verified bit-identity + safety
post-interruption). Two per-frame HashSet<uint> allocations replaced by
reused cleared-in-place fields:
- GameWindow._animatedIdsScratch: WalkEntitiesInto treats null and empty
identically (line 682 `is null || Count == 0`), so an always-non-null
empty set == the old null-when-empty local. Verified.
- RetailPViewRenderer._drawableCellsScratch: flows into
RetailPViewFrameResult.DrawableCells (live ref) but every consumer reads
it same-frame; sigDrawableCells is a per-frame OnRender local that
EmitRenderSignatureIfChanged only FORMATS to a string (no reference
retention, no _lastSig set field). Built frame N -> read frame N ->
cleared frame N+1. No cross-frame corruption.
Build green, full suite 4116 passed / 4 skipped.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
InteriorEntityPartition.Partition allocated a fresh Result (a Dictionary
plus 2 Lists) and a new List<WorldEntity> per newly-seen visible cell,
every frame — the sole call site is RetailPViewRenderer.DrawInside,
once per frame, single-threaded.
Fix: add a Partition(Result, visibleCells, landblockEntries) overload
that clears an existing Result in place (Result.ClearForReuse) and
refills it, reusing each cell's List<WorldEntity> across frames when
the cell key survives (the visible-cell set is normally stable frame to
frame). RetailPViewRenderer now owns one _partitionResult instance
(matching its existing _shellBatch/_buildingGroups/_lookInPrepareScratch
reuse pattern) instead of allocating one per DrawInside call. The
original allocating Partition(visibleCells, landblockEntries) overload
is kept unchanged for tests and one-shot callers (it now delegates to
the reuse overload against a fresh Result).
Bit-identical output required pruning: without removing cell buckets
that end a frame with zero entries, a cell that leaves visibility would
leave a stale empty List sitting in ByCell, changing ByCell.Count and
.Keys enumeration versus the old always-fresh-Dictionary behavior (two
GameWindow diagnostics - sigSceneParticles and FormatPartitionCounts -
read those directly, not just via TryGetValue). Result.PruneEmptyCellBuckets
removes any zero-count bucket after each Partition call, keeping
ByCell's key set exactly what a fresh dictionary would have held.
Verified safe: the sole production consumer (RetailPViewRenderer.
DrawInside) and every downstream reader (WbDrawDispatcher walks,
GameWindow diagnostics) consume partition.ByCell/OutdoorStatic/Dynamics
synchronously within the same frame that built them - no cached
reference survives into the next frame's Partition() call (the one
GameWindow field that stores the reference, _interiorPartition, is
write-only, never read).
dotnet build clean, dotnet test 4120 passed / 4 skipped (unchanged).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
ParticleRenderer.Draw is called up to ~11 times per frame (sky pre/post,
scene, per-visible-cell, dynamics, unattached passes). Each call
allocated a fresh List<ParticleDraw> (BuildDrawList) and a fresh
List<ParticleInstance> (the per-batch `run` list), and ParticleSystem.
EnumerateLive was a `yield return` iterator block - a heap-allocated
state machine allocated fresh on every Draw call regardless of particle
count.
Fix: reuse _drawListScratch/_runScratch fields (Clear() + refill) on
ParticleRenderer. Replace EnumerateLive's iterator with a struct
enumerable/enumerator pair (ParticleSystem.LiveParticleEnumerable): the
foreach fast path uses the struct enumerator directly (zero allocation),
while LINQ/test callers (.ToList(), .Single(), etc.) still work via the
explicit IEnumerable<T> interface implementation, which falls back to a
boxed iterator only when that surface is used - those call sites are
test-only, not the per-frame render path this task targets.
Verified safe: all 11 Draw() call sites in GameWindow.cs are plain
synchronous invocations from the single-threaded OnRender chain (no
Task.Run/Parallel dispatch), and each call fully drains its lists
before returning - no call ever overlaps another's use of the reused
buffers. Per-cell filtering semantics unchanged (same predicate, same
traversal order).
dotnet build clean, dotnet test 4120 passed / 4 skipped (unchanged).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
AnimationSequencer.Advance() allocated a fresh PartTransform[partCount]
every call (BuildBlendedFrame/BuildIdentityFrame), and GameWindow.
TickAnimations reassigned entity.MeshRefs to a fresh List<MeshRef>(partCount)
every tick for every animated entity - including idle NPCs on a breathe
cycle. Both fire every frame in steady state and both become garbage
immediately after being consumed.
Fix: cache a PartTransform[] sized once in the AnimationSequencer
constructor (_setup is readonly and never reassigned, so partCount is
fixed for the sequencer's lifetime) and overwrite it in place each
Advance() call. Cache a List<MeshRef> on the long-lived AnimatedEntity
record (one per animated entity, held in GameWindow._animatedEntities)
and Clear()+refill it each tick instead of allocating a new list.
Verified safe: TickAnimations runs single-threaded to completion before
any draw-side consumer reads MeshRefs (WbDrawDispatcher re-reads
entity.MeshRefs fresh every frame, keyed by entity.Id - it never caches
the list reference across frames). The only place that copies MeshRefs
across an animation boundary (RefreshPaperdollDoll) takes an explicit
`new List<MeshRef>(pe.MeshRefs)` value copy; MeshRef is a readonly
record struct so that copy is unaffected by later in-place mutation of
the source list. No test holds two Advance() results simultaneously.
dotnet build clean, dotnet test 4120 passed / 4 skipped (unchanged).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
User goal clarified: the wildly fluctuating FPS/ms IS the target. Root cause is per-frame throwaway allocations triggering periodic gen2 GC pauses = the spikes. Plan takes only the easy, bit-identical, non-faithfulness-sensitive buffer-reuse sites from the 54-site audit (animation pose buffers, particle draw-lists/iterator, interior partition, trivial per-frame HashSets); defers the two risky sites (EnvCell settled-camera gate, physics Transition pooling). Gate = the existing frame profiler before/after (alloc_kb down, cpu_ms max/p99 tighten) + identical visuals. Loading/pak work parked.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
54-site adversarial allocation audit: town GC churn (1.5-3 MB/frame) is ~90-95 pct OUTSIDE the MP3 draw-submission surface (visibility compute, EnvCell rebuild, particle draw-lists, per-entity anim pose) and independent of the pak. So MP3 will NOT fix the steady-state stutter, and the triage is separable + mostly easy faithfulness-neutral buffer reuse. New MP-Alloc phase pulled ahead of MP1c (load-time smoothness is the rarer hitch). Also flips MP1b to blocked-on-dedup and records the pview-runs-outdoors scenario. Findings in project_mp_track_findings.md.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Session handoff for the next pickup. #176 (purple seam flash) + #177
(stair-room light pop-in) are ONE bug, root-caused to the camera-nearest
MaxGlobalLights=128 snapshot cap evicting in-range lights of visible
cells (Hub = 366 fixtures). Fix deferred to A7 because uncapping exposes
unported per-cell light-reach (through-floor light) + the fixture
falloff-curve misassignment + an unattributed striped-floor artifact.
- New handoff: docs/research/2026-07-06-176-177-handoff-A7-lighting.md
(root cause, why-deferred, the A7 fix order, tooling inventory).
- Roadmap A7 updated: now owns #176/#177; the 'light visibility culling'
hypothesis layer is CONFIRMED (no per-cell insert_light registration);
fixture-curve layer added. Analysis pre-paid.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Full-bake gate baked all 729,888 EnvCells as separate fileId-keyed blobs = 865 GB pak in 52 min, 97 pct near-duplicate cell geometry. The prior 'v1-acceptable duplication' call was wrong by ~100x. Bake-time dedup (compute geomId, extract each unique geometry once, alias all sharing cell-fileIds to one blob offset) is now a v1 requirement; needs no format or reader change and collapses size AND bake time together. MP1b not shipped until the dedup slice lands.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The MaxGlobalLights 128->1024 fix (4d25e04d) was live-tested and made
the eviction pops stop — but with the full 366-fixture pool active,
three unported retail lighting semantics dominate the Facility Hub:
(a) lights reach THROUGH solid floors/walls: retail registers lights
per-CELL (insert_light 0x0054d1b0) so the under-room portals'
purple light never touches the corridor above; our flat
sphere-overlap selection has no reach/occlusion notion — rooms
washed magenta (user screenshot).
(b) stationary weenie fixtures ride the DYNAMIC 1/d falloff (~9x
retail's static 1/d3 bake curve at 3m) — the #143 isDynamic
assignment is wrong for ACE-served world fixtures.
(c) an unexplained striped z-fight-like artifact on lit floor regions
(user screenshot; no coincident dat geometry — the coplanar-pair
sweep came back empty; not a striped texture — all corridor
surfaces are plain Base1Image stone).
Reverted to 128. The cap is now documented as a LOAD-BEARING STOPGAP:
it accidentally approximates per-cell reach by keeping the pool local
to the camera. The #176/#177 root cause (cap eviction popping per-cell
light sets) stays CONFIRMED and fully documented; the real fix is the
A7 dungeon-lighting arc: per-cell light registration + the static
fixture curve + the stripe hunt, THEN uncap. The desired-end-state pin
is kept as Skip with the full pointer. Register row AP-85 rewritten to
match reality; ISSUES #176/#177 back to OPEN with the complete
mechanism story.
Suites: Core 2591+3skip / App 719 green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adversarially-verified review findings 7 and 8:
(7) The DatCollection->IDatReaderWriter adapter existed as THREE
near-identical copies (App-internal original, Bake's copy, Content.
Tests' copy) — a structure where adapter drift is exactly what the
live-vs-pak equivalence suite cannot detect (both sides would only
drift together if they shared one implementation). Now ONE public
AcDream.Content.DatCollectionAdapter next to IDatReaderWriter (GL-free
home established in MP1a), carrying App's FULL behavior including the
[dat-miss] TryGet tripwire log (which now also covers the bake tool
and the equivalence suite) and the caching/locking. All three copies
deleted; WbMeshAdapter (App), BakeRunner (Bake), and
PakEquivalenceTests (Content.Tests) resolve the shared class.
Iteration properties return the REAL dat iterations — the App copy's
hardcoded 0 was a stub nothing read; the unification intentionally
keeps truth (noted in the doc comment). Verified post-move: no
Silk.NET anywhere in Content / Bake / Content.Tests / Bake.Tests
resolved dependency graphs.
(8) Two test gaps closed in PakRoundTripTests: (a) direct on-disk TOC
sortedness — blobs added in DESCENDING key order, then the raw file
bytes parsed (not through the reader) and every TOC entry asserted
strictly ascending; (b) corrupt-blob logging — five repeated reads
through both public paths (TryReadObjectMeshData + ContainsKey) with
stderr captured, asserting exactly ONE [pak-corrupt] line for the
victim key.
Full suite: 4120 tests, 0 failures (Content.Tests 56, Bake.Tests 1,
plus the pre-existing 4 skips).
Adversarially-verified review findings 1 and 10:
(1) The bake pipeline no longer accumulates every decoded ObjectMeshData
in one ConcurrentBag before writing (multi-GB OOM risk on the full
bake), and no longer writes blobs in thread-completion order (which
violated the plan's "bakes must be byte-reproducible run-to-run").
New shape in BakeRunner: build the FULL id list, sort by PakKey, chunk
into 512-id batches; Parallel.ForEach WITHIN each batch; sort each
batch's results by key and AddBlob sequentially; release the batch.
Batches are contiguous key ranges, so the blob region lands in global
key order regardless of thread scheduling, and peak memory is one
batch's output. Side-staged particle-preload meshes drain per batch
into a key-deduped map (first instance wins — per-id extraction output
is deterministic, so instance choice cannot affect bytes) and are
written after all batches, sorted by key, skipping keys already
written.
Program.cs is now a thin arg-parsing shell over the public BakeRunner
so the new dat-gated byte-reproducibility test can drive the REAL
pipeline: tests/AcDream.Bake.Tests (new project, rule 6; registered in
slnx; no Silk.NET in its resolved dependency graph — verified) bakes
the same 9-id mixed fixture twice with DIFFERENT thread counts (8 vs
3 — thread scheduling was the nondeterminism source) and asserts the
two pak files are byte-identical. Ran for real against the dats on
this machine: green.
(10) The isSetup argument for EnvCell extraction now matches at both
call sites (BakeRunner and PakEquivalenceTests both pass false) and is
documented at each: the runtime's own request sites
(WbMeshAdapter.IncrementRefCount / EnsureLoaded) pass isSetup: false
for every MeshRef id including cell-geometry ids. The parameter is
currently dead in MeshExtractor.PrepareMeshData (dispatch is on the
resolved dat type), but two disagreeing call sites were a latent trap.
Header FormatVersion is no longer set by the bake (PakWriter stamps it
per review finding 2, previous commit).
Adversarially-verified review findings 2,3,4,5,6,9:
(2) PakFormat.CurrentFormatVersion=1 constant; PakWriter stamps it
unconditionally (caller header templates can no longer produce a
version-0 pak — the default-0 footgun); PakReader refuses any other
version with a message naming found/expected. Tests: versions 0 and 2
both rejected; writer stamps even when the template omits the field.
(3) Reader robustness, all under the documented corrupt-=-missing
CONTRACT (external-file input — surface loudly once, then behave as
absent; never garbage, never throw from a lookup): (a) per-TOC-entry
bounds validation at open (offset/length outside [header, toc) or
past EOF -> logged once, entry missing, siblings unaffected); (b)
TryReadObjectMeshData catches deserialization failures (malformed
structure behind a matching CRC) -> logged once per entry, false; (c)
structurally unopenable files throw at OPEN with a clear message:
unfinalized header (tocOffset < header size — the placeholder-header
crash signature) and TOC-past-EOF truncation. Tests: corrupt TOC
entry, truncated pak, half-written pak, malformed-blob-behind-valid-
CRC (tamper + CRC recompute) — each verifying sibling blobs still read.
(4) Single-pass read: TryReadObjectMeshData now does ONE ReadArray out
of the map; CRC and deserialization run over the same buffer (was: a
separate VerifyCrc traversal + a second ReadArray + a ToArray copy
inside Serializer.Read). New Serializer.Read(byte[]) overload avoids
the defensive copy. Verdict set stays a ConcurrentDictionary (MP1c
calls this from 4 decode workers). True span-over-mmap zero-copy is
deferred to MP1c profiling per the class doc comment.
(5) PakWriter.Dispose restored to try/finally: Finish() does real I/O
and can throw (disk full) — the stream must ALWAYS close so no file
handle leaks mid-unwind. (Reverts the a5926ebc simplification, which
was wrong about this.) Test: dispose after an AddBlob exception leaves
the file deletable.
(6) CRC-32 known-answer vectors: "123456789" -> 0xCBF43926, empty ->
0x00000000, single zero byte -> 0xD202EF8D. The suite was previously
blind to a self-consistent-but-wrong CRC.
(9) Equality comparator floats switched to bit-equality
(SingleToInt32Bits/DoubleToInt64Bits, incl. Vector2/3 + Matrix4x4
components): for byte-identity round-trip purposes `==` was both too
strict (NaN==NaN false — a surviving NaN payload would wrongly FAIL)
and too lax (-0.0==+0.0 true — a sign-bit flip would wrongly PASS).
Content.Tests: 54/54 green (was 43).
MP1b review finding, verified against EnvCellRenderer.GetEnvCellGeomId:
runtime dedups interior geometry under a 64-bit content hash (bit 33),
not the cell fileId the pak keys by. MP1c maps at the RegisterCell seam
(read pak[cellFileId], register under geomId, dedup post-load as today);
duplicate blob content on disk is v1-acceptable. GeomId-keyed baking is
not viable in format v1 (hash exceeds PakKey's 56 usable bits).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The probe launch discriminated it: the user reproduced the purple floor
flash while [light] (ambient branch) and [pv-input] (portal flood) read
provably healthy — eliminating the last CPU-side theories and exposing
the one channel the probes could not see: per-cell 8-light set
composition.
BuildPointLightSnapshot kept the MaxGlobalLights=128 point lights
nearest THE CAMERA; the Facility Hub registers 366 fixtures, so 238
were evicted per frame by camera distance. SelectForObject (faithfully
camera-independent, and unit-pinned as such) could only choose from the
surviving 128 — an in-range torch of a visible cell that ranked past
the cap dropped out of that cell's 8-set, so per-cell Gouraud lighting
flipped as the chase boom swung the camera:
- #176: the flipping unit is a CELL -> discontinuity lines at exactly
cell-seam granularity; a torch-losing floor drops to dim blue-grey
stone at 0.2 ambient (the perceived purple), camera-angle dependent.
- #177: a stair room whose torches all ranked past the cap rendered at
bare 0.2 ambient (near-black = 'not visible'); approach re-admitted
them ('pops into existence'); the sweeping boundary dropped the
ramp's lights mid-descent ('disappears on the last step'). The
geometry never vanished - its lights did.
Retail's minimize_object_lighting (0x0054d480) has NO global
camera-nearest pool cap (lights register per cell, insert_light
0x0054d1b0). Fix: MaxGlobalLights 128 -> 1024, a non-biting safety
valve (GlobalLightPacker grows to fit; 64 B/light). Register row AP-85.
TDD pin: PointSnapshot_HubScaleLightCount_ObjectSelectionIsCameraInvariant
(RED at 128 with a Hub-scale 401-light layout, GREEN at 1024). The
pre-existing camera-independence pin covered the SELECTOR but not the
SNAPSHOT it selects from - the pop re-entered one stage upstream.
Suites: Core 2588 / App 719 / UI 425 / Net 385 green. Pending user gate.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Self-review finding: Dispose()'s best-effort Finish() call was wrapped
in a try/catch for InvalidOperationException that could never fire —
the surrounding `if (!_finished)` guard already excludes the only
condition under which Finish() throws that exception. Dead defensive
code that LOOKS like it might be swallowing a real failure is exactly
the kind of thing the no-workarounds policy asks to avoid; removed with
a comment explaining why no catch is needed. Behavior unchanged; 43/43
Content.Tests still green.
PakEquivalenceTests runs MeshExtractor LIVE and bakes the SAME fixture
ids to a temp pak, reads it back via PakReader, and deep-compares every
field via the Task 3 ObjectMeshDataEquality comparator. Since the bake
tool and the live client drive the identical MeshExtractor code (MP1a),
this proves the pak ROUND-TRIP preserves what extraction actually
produces on real content — the serializer's job, not a re-verification
of the extraction algorithm itself (the existing Conformance suite owns
that).
Fixture set (>= the plan's minimums): 10 GfxObjs (3 known-tricky ids
reused from Issue119UpNullGfxObjDumpTests - 0x010002B4, 0x010008A8,
0x010014C3 - plus 7 more from dat order), 3 Setups reused from door/
tower conformance fixtures (0x020019FF, 0x020005D8, 0x020003F2), 5
EnvCells walked from the Holtburg landblock 0xA9B40000's
LandBlockInfo.NumCells range (same idiom as
StipplingSurfaceEquivalenceTests).
Skips cleanly when dats are absent via ContentConformanceDats.
ResolveDatDir(), a deliberate small duplicate of
ConformanceDats.ResolveDatDir()'s exact pattern (env var then
Documents/Asheron's Call fallback) since Content.Tests cannot reference
the AcDream.Core.Tests project. ContentTestDatCollectionAdapter is a
third small copy of the IDatReaderWriter dat-access glue (alongside
AcDream.App's internal original and AcDream.Bake's copy) for the same
layering reason (structure rule 6: tests live in the project matching
the layer under test).
Ran for real against the dats on this machine (not just CI-skip path):
1 test, all fixture ids extracted live with zero failures, baked,
read back, and field-for-field identical. Full solution dotnet test:
4106 tests total (43 Content.Tests + 385 Core.Net.Tests + 425
UI.Abstractions.Tests + 722 App.Tests incl. 2 pre-existing skips + 2531
Core.Tests incl. 2 pre-existing skips) — 0 failures.
New offline console tool driving AcDream.Content.MeshExtractor (the SAME
extraction code the live client runs, per MP1a) to produce a versioned
pak file. Arguments: --dat-dir (required), --out (default acdream.pak
next to the dats), --ids/--landblocks (dev filters), --threads (default
Environment.ProcessorCount).
Enumeration mirrors existing idioms rather than inventing new ones:
GfxObj/Setup ids via DatCollection.GetAllIdsOfType<T>() (src/AcDream.Cli/
Program.cs); EnvCell ids by walking dats.Cell.Tree for LandBlockInfo
entries (low 16 bits == 0xFFFE) and then the firstCellId+NumCells range
per landblock (the same idiom as GameWindow.BuildPhysicsDatBundle /
BuildInteriorEntitiesForStreaming). GetAllIdsOfType<T>() does not cover
cell-dat range-based types, hence the manual walk.
BakeDatCollectionAdapter is a from-scratch copy of AcDream.App.Rendering.
Wb.DatCollectionAdapter (which is `internal` to AcDream.App and, more
importantly, referencing AcDream.App would drag Silk.NET/GL into the
bake tool — a hard violation of "no Silk.NET anywhere"). It's plain dat-
access glue, not an AC-specific algorithm, so a small duplicate is the
right call over inventing a shared-but-App-rooted package.
Particle-preload GfxObjs MeshExtractor side-stages mid-extraction
(sideStagedSink, thread-safe ConcurrentQueue per MP1a's documented
contract) are deduped against the primary GfxObj set and against each
other before being written as their own GfxObjMesh entries.
ConsoleErrorLogger is a minimal hand-rolled ILogger (stderr, Warning+)
rather than NullLogger — MeshExtractor's LogError/LogWarning calls on
malformed dat entries must stay visible per the project's "logger
injection for silent catches" lesson.
Progress line every 5s (baked/total, failures, elapsed, ETA); a
malformed dat entry is caught per-id and counted as a failure, never
fatal — matches the runtime's own per-id try/catch behavior. Failures
report is printed at the end, capped at 200 lines.
Smoke-tested against the real dats on this machine: --ids
0x01000001,0x02000001 baked 1 GfxObj (40 vertices) + 1 Setup (34 parts)
in 1.4s with 0 failures; a follow-up run adding EnvCell 0xA9B40100
(Holtburg's first interior cell) baked all three asset types
successfully. PakReader opened both scratch paks and read back
deep-correct ObjectMeshData; scratch files deleted after verification.
*.pak added to .gitignore in this commit per the plan.
TDD: PakRoundTripTests written first, confirmed a compile failure
against the not-yet-existing PakWriter/PakReader types.
PakWriter streams [header placeholder][64-byte-aligned blobs][TOC sorted
by key], then seeks back and finalizes the header once TocOffset/TocCount
are known (plan: "TOC last" so the writer doesn't need blob count
up front). PakReader mmaps the whole file once, loads the TOC into a
sorted array for O(log n) binary-search lookup, and verifies each blob's
CRC-32 LAZILY on first access (cached per-index so a corrupt blob logs
exactly once and is thereafter always treated as missing rather than
handing back garbage bytes). No locks anywhere — the mapped memory is
immutable for the reader's lifetime by construction.
CRC-32 implemented directly (standard IEEE 802.3 table-driven variant)
rather than adding a System.IO.Hashing package reference, keeping
AcDream.Content's dependency surface minimal per its existing
"no GL binaries" intent.
9 tests green: header round-trip, every-key-found, blob deep-equality
via the Task 3 comparator, missing-key returns false, corrupted-blob
(single flipped byte) is detected and treated as missing while leaving
sibling blobs unaffected, 64-byte blob alignment, and TOC binary search
cross-checked against a linear scan for both present and absent keys.
The dungeon render pair (purple seam flash + stair pop-in) resisted
desk root-causing, but the investigation narrowed the space to two
live theories and shipped permanent apparatus:
- Issue176177DungeonSeamInspectionTests: dat truth — corridor floors
ARE textured drawn PortalSide portal polys; the reciprocal is NoPos;
the 'stairs' are 0x8A020182's ramp shell (vertical portals, zero
statics); CellBSP partitions exactly at portal planes; DXT1 textures
carry zero transparent-mode texels.
- Issue176177FacilityHubFloodReplayTests: production-matched flood
replays — approach/descent/gaze-sweep/walk all healthy with coherent
inputs; ScenarioE pins the flood's collapse-to-root sensitivity under
incoherent (root, eye) pairs.
- Issue176177SeamTransitLagTests: the resolver flips cells within one
tick-step of the portal plane — the 0.33-0.47m [cell-transit] 'lag'
in the gate logs is speed*tick quantization, not membership error.
Refuted (do NOT retry — ledger in the research doc): placeholder
texture, reciprocal z-fight, seal z-fight (seals only fire for
OtherCellId==0xFFFF), root/eye incoherence (production camera sweep is
mm-exact at planes), flood bistability, #119-class statics, undefined
DXT mips (both paths decode DXT->RGBA8; the compressed-array branch is
dead code), DXT1 alpha, fog mix (ramp ~538m), lightning leak (flash==0
in production), viewer-light pops (smooth (1-d/range) ramp).
Filed #178 (A8 double-sided shell stopgap still live) and #179
(lightning flash lacks an indoor gate — dormant). The purple can only
be the fog clear color (undrawn pixels) or the outdoor ambient+sun
tint; discrimination needs ONE probe launch (ACDREAM_PROBE_LIGHT +
ACDREAM_PROBE_PVINPUT + ACDREAM_PROBE_CELL) — protocol in
docs/research/2026-07-06-176-177-render-pair-investigation.md.
Suites: Core 2587 / App 719 green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
TDD: ObjectMeshDataSerializerTests + the shared ObjectMeshDataEquality
field-by-field comparator (reused by Task 6's equivalence suite) written
first, confirmed a compile failure against the not-yet-existing
ObjectMeshDataSerializer type.
Serializes EVERY field of the ObjectMeshData family per the plan's
normative layout: primitives raw LE, arrays as count:i32+payload,
blittable arrays (VertexPositionNormalTexture[], ushort[], byte[]) via
MemoryMarshal.AsBytes bulk copy, TextureBatches written sorted by the
(Width, Height, Format) key tuple for run-to-run determinism regardless
of dictionary insertion order, nullable fields as present:byte+value.
EnvCellGeometry nests recursively (MeshExtractor can populate one level
today; the serializer supports arbitrary depth rather than assuming it).
Namespace-trap finding: StagedEmitter.Emitter resolves to
DatReaderWriter.DBObjs.ParticleEmitter (the dat DBObj, verified via
reflection against the pinned Chorizite.DatReaderWriter 2.1.7 package
and confirmed live by MeshExtractor's `emitter.HwGfxObjId.DataId` call
site compiling), NOT AcDream.Core.Vfx.ParticleEmitter (the runtime
particle-simulation type with a live Particle[] pool that would NOT be
serializable asset data). All ~31 of its fields are written explicitly
rather than delegating to its own Pack/Unpack, which require a live
DatBinWriter/DatBinReader bound to a DatDatabase — coupling our pak's
determinism to a third-party wire-format helper we don't control the
versioning of.
33 tests green: 9 round-trip fixtures (empty/vertices+indices/multi
texture-batch-groups/setup-parts/emitters/nullable-present/nullable-
absent/edge-lines/nested-EnvCellGeometry), same-instance-twice byte-
identity, and dictionary-insertion-order-independence (two orders ->
identical bytes) plus a key-sort-order assertion on the raw bytes.
TDD: PakKeyTests + PakFormatTests written first (confirmed a compile
failure against the not-yet-existing AcDream.Content.Pak namespace),
then PakKey (64-bit type:u8|fileId:u32|reserved:u24 compose/decompose)
and PakFormat (64-byte PakHeader, 24-byte PakTocEntry) implemented to
the normative layout in the MP1b plan. 21 tests green, including a key-
ordering test proving ascending numeric key order equals ascending
(type, fileId) tuple order (the TOC binary-search precondition) and an
explicit byte-offset test for both structs.
New xunit test project for the Content layer (rule 6: tests live in the
project matching the layer under test). ProjectReference to
AcDream.Content; TreatWarningsAsErrors + LangVersion latest per MP1b's
csproj requirements. Registered in AcDream.slnx. Placeholder PakKeyTests
smoke test becomes the real Task 2 suite next.
Normative v1 format (ACPK header with dat-iteration stamps, sorted
24-byte TOC entries with crc32, 64-byte-aligned uncompressed blobs,
deterministic ObjectMeshData serialization), the bake CLI that drives
MP1a MeshExtractor in parallel, the mmap zero-copy PakReader, and a
dat-gated live-vs-pak equivalence suite. Scope-fenced to the
ObjectMeshData asset classes (the decode-storm content from the MP0
baseline); terrain/BSP/scenery/degrade blobs are later slices. Also
amends the spec: PakReader lives in AcDream.Content (Content->Core
direction from MP1a makes a Core home circular).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Launch smoke passed: world renders identically (user-confirmed), zero
dat-layer tripwires in the session log, frame profile identical to the
pre-extraction baseline (town p50 2.1-2.2ms, same GC cadence) - the
move was perf-neutral as a pure relocation should be. Roadmap Track MP
table updated: MP1 split into shipped MP1a and upcoming MP1b (pak +
bake tool + reader) / MP1c (streaming cutover + hitch gate).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Coordinator-directed final cleanup before the user gate; none behavioral:
1. MeshExtractor public surface narrowed to the cross-assembly entry
points App actually calls (PrepareMeshData, PrepareCellStructMeshData,
CollectParts, ComputeBounds); PrepareSetupMeshData,
CollectEmittersFromScript, PrepareGfxObjMeshData,
PrepareEnvCellMeshData, PrepareCellStructEdgeLineData back to private
(internal dispatch, only reached via PrepareMeshData).
2. sideStagedSink constructor parameter is now REQUIRED (no default;
type stays nullable for a conscious null): a bake tool that forgot
the sink would silently lose particle-preload meshes.
3. AcDream.Content.csproj gains TreatWarningsAsErrors + LangVersion
latest (parity with AcDream.Core.csproj). Surfaced zero warnings.
4. Dead usings removed from ObjectMeshManager.cs (BCnEncoder.*,
SixLabors.*) — the inline decode moved out in Task 4.
5. Doc fixes: ObjectMeshData.cs cross-assembly <see cref> ->
plain text (Content can't resolve App types); IDatReaderWriter.cs
stale Phase O-T7 'both in this namespace' sentence rewritten.
6. Stale test doc comments updated to MeshExtractor.PrepareGfxObjMeshData
(StipplingSurfaceEquivalenceTests, Issue119UpNullGfxObjDumpTests) —
comments only, no code/assertion changes.
dotnet build green (0 warnings in Content under warnings-as-errors);
full test suite 4059 passed / 0 failed / 4 skipped.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Coordinator-directed follow-up. The buffer-and-drain seam diverged from
the original on the exception path: pre-MP1a, CollectEmittersFromScript
enqueued particle-preload meshes DIRECTLY into _stagedMeshData
mid-Prepare, so preloads staged before a later throw in the same
Prepare* call (reachable via PrepareEnvCellMeshData side-staging during
its StaticObjects loop, then PrepareCellStructMeshData throwing on a
malformed-dat texture decode) were already safely enqueued. The drain
version only flushed after a successful return — on throw, entries
stranded on the shared extractor until an unrelated successful call
flushed them, and were silently dropped on dispose.
Fix: MeshExtractor takes an Action<ObjectMeshData>? sideStagedSink
constructor parameter; the two CollectEmittersFromScript sites become
_sideStagedSink?.Invoke(meshData) — the original code shape (immediate
hand-off) at those exact lines. ObjectMeshManager wires the sink to
_stagedMeshData.Enqueue, restoring the original immediate-enqueue
semantics including on mid-Prepare throw. _sideStaged buffer,
DrainSideStaged(), and the ProcessQueueAsync drain loop are deleted.
The MP1b bake tool passes its own collector.
Inventory doc updated: MP1a note now records the sink seam and the
Content-owned upload enums, so its no-behavior-change claim is accurate.
dotnet build green; full test suite 4059 passed / 0 failed / 4 skipped.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
MP1a ground truth: the runtime atlas consumes decoded RGBA8 via
TextureBatchData, so baking to BC7/BC1 (lossy) would change delivered
pixels and contradict the phase's own byte-identical conformance gate.
V1 stores RGBA8 exactly as the runtime produces; BC becomes an explicit
post-conformance option with its own visual gate. Mip chains stay
runtime-generated in v1.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Coordinator-directed follow-up: AcDream.Content must stay Silk.NET-free
(the MP1b bake tool must not ship GL binaries). The Silk.NET.OpenGL
PackageReference added for the PixelFormat?/PixelType? upload hints is
replaced by Content-owned UploadPixelFormat/UploadPixelType enums
(UploadFormats.cs) whose underlying values are the GL ABI constants
(Rgba = 0x1908, UnsignedByte = 0x1401), verified numerically identical
to the Silk.NET.OpenGL members against 2.23.0. This is the one
sanctioned edit to the verbatim-moved Prepare* bodies: enum literal for
enum literal, numeric value identical, behavior unchanged. App casts at
the single upload boundary (AddTexture call in UploadGfxObjMeshData)
via lifted nullable enum conversion — value- and null-preserving.
Also hardens the MP1a _sideStaged hand-off seam: List -> ConcurrentQueue.
One MeshExtractor is shared by up to MaxParallelLoads (4) decode workers;
the original code enqueued to the thread-safe _stagedMeshData directly,
so the hand-off buffer must be thread-safe too. Drain ordering verified:
side-staged entries enqueue BEFORE the top-level result, preserving the
original mid-Prepare FIFO order.
Verified: grep -i silk on the csproj -> no matches; deps.json has zero
Silk entries; dotnet build 0 errors; full test suite green (4059 passed).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The corrected 1.835m collision capsule blocks the window alcove like
retail; doorways, seams, and stairs unaffected by the taller head.
Remaining #137 scope: the doors half (block/pass per open state, the
#175 pose fix's pending gate).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Mechanical-move plan: lift TextureKey out of the GL atlas class, move the
ObjectMeshData family, extract the Prepare* CPU pipeline verbatim into a
new MeshExtractor in AcDream.Content (net10.0, no Silk.NET), App keeps
queue/worker/upload lifecycle and delegates. Binding rules: verbatim
bodies (no SurfaceDecoder switch - byte-identity is MP1b conformance
foundation), compiler enumerates dependencies, BLOCKED on any hidden GL
dependency, zero divergence rows. Gates: full suite green per task +
user launch smoke. Boundary facts verified against source this session:
the Prepare* region touches GL only via the TextureKey struct; first GL
call is UploadGfxObjMeshData.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The 'run into the last corridor window and pop up through its roof'
report: the live callers passed sphereHeight: 1.2f into
SpherePath.InitPath, whose head-sphere formula (height - radius) put the
head sphere center at 0.72 - the capsule top at 1.2m. The top 0.63m of a
1.83m character had NO collision, so at the corridor-end window alcove
(0x8A020179 -> 0x8A02017E: 0.70m sill face, 1.3m opening, sloped funnel
behind) the step-up's placement never saw the head overlapping the
lintel solids and let the player climb in head-through-roof.
Dat truth (HumanSetup_CollisionSpheres_DatTruth): Setup 0x02000001
spheres = (0,0,0.475) r=0.48 and (0,0,1.350) r=0.48 - capsule top 1.83 =
Setup.Height 1.835. Retail collides with that sphere list verbatim
(CPhysicsObj::transition 0x00512dc0 -> init_sphere(GetNumSphere,
GetSphere, m_scale)).
Fix: PlayerMovementController + the GameWindow remote resolve now pass
sphereHeight: 1.835f (capsule top; head center 1.355 vs dat 1.350).
InitPath unchanged - captured-input replay fixtures (recorded 1.2
inputs) stay byte-identical. Register TS-46: the (radius, capsule-top)
scalar approximation of the Setup sphere list (5mm foot/head offsets;
remotes use human dims) with the retire path (plumb the sphere list).
Pins: WindowOpening_HeadCannotFit_EntryBlocked (22-frame walked approach
wall-slides at the sill, never enters 0x8A02017E) +
WindowAlcove_RaisedPlacement_HeadInLintelSolid_Collides (Path-1
placement rejects the raised head in the lintel solids) + the
WindowShaft_FullPolyDump / HumanSetup dat inspections.
Suites: Core 2562 / App 713 / UI 425 / Net 385, 0 failures.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Full-route capture (Holtburg -> dungeon 0x0007 -> town 0xCE94 -> dungeon
-> Holtburg, 4 teleports) plus the supplemental town run. Verdict: the
render-side-CPU split is confirmed (GPU <= ~2.7ms, upd/upl/imgui ~= 0),
steady medians beat the spec assumption (worst town p50 3.6ms), and the
smoothness gap is GC: 1.5-3 MB allocated per frame drives gen2
collections 1-2/s and every town p99/max violation, while dungeon
windows (zero gen1/gen2) are spike-free at ~2000 fps. Teleport hitch
quantified at 211ms worst frame / 75.7MB single-frame allocation.
Gate decision: PROCEED to MP1 unchanged, with one recorded amendment -
a bounded post-MP1 allocation-triage session for churn sites outside
the MP3 rewrite surface. Fort Tethana axiom view still to be
re-measured at the MP2/MP3 gates.
MP0 complete: profiler shipped (7d74c68c..4b44a152) + baseline + gate.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
User gate 2026-07-06 evening: 'not collision anymore. Good.' — the
corridor-phantom arc (stub sliding-normal leak, slide_sphere opposing
branch, CheckOtherCells stale footCenter) is user-verified fixed.
Same gate surfaced three follow-ups:
- #137 remaining scope: the window/opening climb — running into the last
corridor window steps the player up into the opening, head+camera
through its roof. Log anchor: transit 0x8A020179 -> 0x8A02017E at
(90.209,-41.774,-5.209) — sill ~0.8m > stepUpHeight 0.6, climbed anyway.
- #176 (NEW, render): purple floor flashing at cell seams, camera-angle
dependent — survives the physics fix, so render-side; the seam floors
ARE portal polygons to the under-rooms (placeholder-surface suspicion).
- #177 (NEW, render): stairs between levels pop in/out — invisible from
the corridor, appear on entering the room, vanish on the last step
down. The #119 visibility class, dungeon edition.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Roadmap gains the Track MP phase table (MP0 in progress, MP1-MP5 ahead);
milestones doc records the side track + the explicit user-authorized
freeze exception (MP may reopen frozen streaming/WB-rendering subsystems
under its own gated phases only); CLAUDE.md Current state gets the
one-line pointer.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Three review finds, all in the runtime-toggle path or edge math (the
steady-state path was confirmed correct):
1. GPU stale-slot across toggle-off/on: the disabled branch now
Disposes the GpuFrameTimer (and nulls it) instead of Stop()-ing and
keeping it — a kept instance would poll slots left pending from
BEFORE the pause on re-enable and report temporally stale GPU
samples. The re-enable branch's existing null guard rebuilds the
ring fresh. Dispose is safe there: the branch runs at the top of
OnRender with the GL context current.
2. Stale stage accumulation across mid-stage toggle-off: a StageScope
disposed after the flag flips still calls EndStage, leaving a
partial delta in _stageAccumTicks. The re-baseline branch now
Array.Clear()s it alongside the GC re-baselining.
3. FrameStatsBuffer.Max() seeded with 0, clamping all-negative windows
(the alloc channel can go negative if the boundary ever crosses
threads) and disagreeing with Percentile(). Now seeds from the
first live sample after the empty check (slots [0.._count) are
always the live window regardless of ring wraparound); empty still
returns 0 to match Percentile. TDD: Max_AllNegative_ReturnsTrueMax
failed (returned 0) before the fix, passes after.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The plan's FormatReport_IsInvariantAndComplete test ends with a
DoesNotContain(",0") guard (after stripping the gc=3/1/0 token) proving
the formatter emits period decimals regardless of the host culture.
It was dropped in the Task 3 commit; restored verbatim. Passes as-is —
FormatReport already uses CultureInfo.InvariantCulture throughout.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The P2 cellar-lip lesson one loop deeper. CheckOtherCells takes footCenter
by value and used it for every cell in the loop — but a mid-loop query can
MOVE the sphere: at the Facility Hub cell boundaries the neighbor's ramp
floor full-hits, step_sphere_up climbs the foot +0.6mm and returns OK, and
the loop continued querying the REMAINING cells (including the under-room,
portal-ring-3) at the pre-climb height — 0.4mm inside the double-faced
floor slab, grazing its underside (the under-room's ceiling) within the
near-miss window. That dispatched a neg-poly step-up with a DOWNWARD
normal, whose failure funneled into slide_sphere's opposing branch ->
synthetic reversed-movement collision -> Collided -> revert, every frame:
the seam shake (and, pre-mechanism-2-fix, the original absorbing wedge's
entry).
Retail check_other_cells reads the LIVE sphere_path.global_sphere for
every cell (each cell's find_collisions receives the transition itself,
pc:272717+). Fix: re-read footCenter = sp.GlobalSphere[0].Origin at the
top of each loop iteration.
All three Issue137CorridorSeamReplayTests repros un-skipped: the
snapshot-exact west-boundary crossing (capture tick 4101), the east
deep-straddle, and the clean-run lifecycle all GREEN. Full suites: Core
2556 / App 713 / UI 425 / Net 385, 0 failures.
Visual gate pending: corridor run + the purple seam flashing re-check
(expected to be the render exposing the same per-frame oscillation).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
GameWindow gets one field (_frameProfiler), one FrameBoundary() call at
the top of OnRender, three stage scopes (Update in OnUpdate, Upload
around _wbMeshAdapter?.Tick(), ImGui around _imguiBootstrap.Render()),
and one Dispose() call in teardown before _dats?.Dispose() releases the
GpuFrameTimer query ring. No new feature bodies land in GameWindow.cs —
all profiler logic lives in AcDream.App.Diagnostics.
DebugVM.FrameProf mirrors RenderingDiagnostics.FrameProfEnabled so the
DebugPanel checkbox ("Frame profiler ([frame-prof])") toggles the 5s
report live. Note: the checkbox lives in
AcDream.UI.Abstractions/Panels/Debug/DebugPanel.cs (IPanelRenderer,
backend-agnostic) alongside the other Diagnostics-section checkboxes —
not in AcDream.UI.ImGui, which holds no panel-drawing code per Code
Structure Rule 3.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Permanent frame profiler: FrameBoundary() at the top of OnRender measures
CPU frame time (swap-to-swap delta), brackets the frame in a GpuFrameTimer
TimeElapsed query (self-disabled under ACDREAM_WB_DIAG=1), and samples
per-frame allocated bytes + GC collection deltas. BeginStage(FrameStage)
scopes attribute CPU time to Update/Upload/ImGui. Emits one [frame-prof]
report line every ~5s while RenderingDiagnostics.FrameProfEnabled is on;
zero-cost stage scopes when off. Report formatting is a pure static
method, unit-tested for invariant-culture formatting and the gpu=off
fallback text.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
RenderingDiagnostics.FrameProfEnabled is the master toggle for the
upcoming permanent frame profiler (ACDREAM_FRAME_PROF=1, runtime
mirror via DebugVM.FrameProf). GpuFrameTimer wraps a depth-4 ring of
TimeElapsed queries for whole-frame GPU time, mirroring
WbDrawDispatcher's query idiom (including the #125 never-begun-slot
guard). Caller must not run this while ACDREAM_WB_DIAG=1 owns
TimeElapsed queries — GL forbids nested TimeElapsed queries.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Fixed-capacity ring buffer of long samples with nearest-rank percentile
and max over the current window. Pure, allocation-free after
construction. Foundation for the MP0 frame profiler
(docs/superpowers/specs/2026-07-05-modern-pipeline-design.md).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The Issue137 replay tests carry the tick-4101 snapshot inline — the raw
jsonl is a session artifact, not a fixture.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
