acdream/docs/research/2026-07-06-176-177-render-pair-investigation.md

#176/#177 render pair — investigation ledger (2026-07-06, session 2)

✅ OUTCOME (same day, after the probe launch): ROOT CAUSE FOUND + FIX SHIPPED

The probe run discriminated it. The user reproduced the purple floor flash while BOTH surviving CPU theories read provably healthy in the log — [light] insideCell/ambient rock-stable (one pre-spawn outdoor line, then flat 0.2 grey through 36 transits), [pv-input] flood stable (54k frames, zero collapses). That eliminated T-A and T-B and exposed the one channel the probes were structurally blind to: per-cell 8-light SET COMPOSITION.

The log's own headline number told the story: registeredLights=366 — against MaxGlobalLights = 128. BuildPointLightSnapshot kept the 128 lights nearest THE CAMERA and evicted 238 every frame; SelectForObject (camera-independent, faithfully retail — 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 → the cell's per-vertex Gouraud lighting flipped as the camera moved (the chase boom swings the camera position by meters, re-ranking the 128):

• #176: the flipping unit is A CELL → discontinuity lines at exactly cell-seam granularity; a torch-losing floor drops to dim blue-grey stone (0.2 ambient × stone = the perceived purple); camera-angle dependent by construction.
• #177: a stair room whose torches ALL ranked past the cap rendered at bare 0.2 ambient — near-black in a dungeon = "not visible from the corridor"; approaching re-admitted them = "pops into existence"; the boundary sweeping during the descent dropped the ramp's lights = "disappears on the last step". The geometry never vanished — its LIGHTS did.

Retail anchor: minimize_object_lighting (0x0054d480) selects from the cell-registered reaching set (insert_light 0x0054d1b0) — no global camera-nearest pool cap exists in retail.

Fix: MaxGlobalLights 128 → 1024 (a non-biting safety valve; the GPU packer grows to fit — 64 B/light). Register row AP-85. TDD pin: LightManagerTests.PointSnapshot_HubScaleLightCount_ObjectSelectionIsCameraInvariant (RED at 128 with a Hub-scale 401-light layout, GREEN at 1024). All four suites green. Pending the user visual gate.

Process note: the pre-existing test SelectForObject_CameraIndependent_DependsOnlyOnObjectCentre was written to pin "the property that kills the lights-up-as-I-approach popping" — it proved the SELECTOR camera-independent while the SNAPSHOT it selects from was camera-capped. The pop re-entered one stage upstream of the pin.

---

Pre-launch status below (kept as the audit trail): mechanism NOT yet pinned — but the hypothesis space is now razor-thin. Twelve candidate mechanisms refuted by direct evidence (dat dumps, headless replays, production-log analysis, code reads). Every layer that can be checked offline is verified HEALTHY at the anchor cells. The surviving discriminator requires ONE live probe launch (protocol at the bottom — piggyback on the pending #175 door gate).

The issues

• #176 — purple flashing on dungeon floors at cell seams, camera-angle dependent (Facility Hub).
• #177 — stairs pop in/out across levels: (a) vanish on the last step running down, (b) invisible looking into the stair room from the corridor, (c) pop into existence on entering.

Anchor-cell dat truth (Issue176177DungeonSeamInspectionTests)

• Corridor 0x8A02016E floor = three abutting TEXTURED drawn portal polys (polys 1/3/5, surface 0x08000377 → DXT1 tex 0x050026F7, [PortalSide]) → under-hall 0x011E (z=−12). Reciprocal ceiling poly = NoPos (not drawn).
• The "stairs" = a RAMP owned by 0x8A020182 (inclined drawn polys, z −9…−6 floor + −6…−3 ceiling); 0x0183 = flat lower cell. Connections are VERTICAL wall portals (NOT floor-portals). PortalSide flag asymmetry: 0x0182→0x0183 carries PortalSide; the back-portal does not.
• Zero StaticObjects in all five anchor cells (no #119-class statics, no torch-bearing stabs → no registered point lights except the viewer fill).
• CellBSP volumes partition EXACTLY at the portal planes (no overlap zone).
• All surfaces resolve; DXT1 textures contain zero transparent-mode texels (all 3-color-mode blocks, index 3 never used).

REFUTED mechanisms (each by direct evidence — do NOT retry)

#	Hypothesis	Killed by
1	Placeholder/missing texture (magenta class)	All surfaces resolve; drawn-poly sweep 0 misses
2	Reciprocal portal-poly z-fight	Reciprocal is NoPos (never drawn)
3	Seal depth-stamp z-fights the drawn floor-portal	Seals fire ONLY for OtherCellId==0xFFFF (GameWindow:11437); a sealed dungeon draws zero seals
4	Root/eye incoherence (viewer root lags the eye across portal planes)	Production camera sweep publishes coherent pairs; out-cell flips at x=85.001/88.335 — mm-exact at the planes (gate2 log). The flood DOES collapse to 1 cell under artificially incoherent inputs (ScenarioE pin) — but production inputs are coherent
5	Membership transit lag (0.33–0.47 m in [cell-transit]) as the render-root lag	The resolver flips within one tick-step of the plane in the harness (Issue176177SeamTransitLagTests); the logged "lag" is speed×tick quantization of the PLAYER probe, and the camera root (probe out-cell) is plane-exact
6	Flood bistability at the anchors	ScenarioC gaze sweep (2° steps, 4 pitches): 0 one-step drops; ScenarioA stair approach: ramp+lower admitted at all tested eyes/pitches
7	Staircase = EnvCell static culled by viewcone (#119 class)	Zero statics in the anchor cells; the stairs are shell geometry
8	Undefined DXT mip levels (compressed arrays skip GenerateMipmap)	Both ObjectMeshManager texture paths DECODE DXT→RGBA8 (BcDecoder) — the compressed-array branch of ManagedGLTextureArray is dead WB-heritage code; RGBA8 arrays get real mips
9	DXT1 3-color-mode alpha=0 texels + opaque-pass discard a<0.05 / A2C	Block histogram: 0 transparent texels in all Hub floor/wall textures
10	Fog mix toward purple FogColor at distance	Fog ramp starts at _nearRadius×192×0.7 ≈ 538 m (radii stay 4/12 in dungeon mode); Hub sightlines ≤ ~100 m → fog term ≡ 0
11	Lightning-flash additive (uFogParams.z × (0.6,0.6,0.75)) leaking indoors	WeatherState._flashLevel is 0 in production ("Production never TriggerFlashes") — dormant test hook. (The missing indoor gate is still real debt for when storm strobes ship.)
12	Viewer-light per-cell/per-vertex pops (hard range edge or 8-set membership flips)	The point ramp is (1−d/range) — smoothly ZERO at the range boundary; set-membership beyond range is a zero-contribution no-op. No torches exist in the Hub cells to churn the 8-cap

VERIFIED-HEALTHY layers (offline pins, keep as regression assets)

• PortalVisibilityBuilder at the anchors: approach/descent/gaze-sweep/walk scenarios (Issue176177FacilityHubFloodReplayTests) — admissions correct and stable with coherent inputs; the ScenarioE incoherent-input collapse is the sensitivity pin (1-cell flood when root≠eye side).
• Membership: ResolveWithTransition flips cells within one tick of the portal plane, both directions (Issue176177SeamTransitLagTests).
• Mesh path: CellMesh.Build (production, GameWindow:7013) draws the textured floor-portal strips; snapshot frustum gate (WbFrustum) is the standard conservative p-vertex test; per-cell AABBs are vertex-derived, 8-corner transformed.
• Per-instance light sets are truly per-instance (EnvCellRenderer MDI).

SURVIVING theories (need the live discriminator)

• T-A (ambient flip): any frame where playerRoot resolves null (or playerSeenOutside defaults true) runs the OUTDOOR lighting branch — purple sky ambient + full sun. Sun is directional-from-above → floors (N·L≈1) catch it, walls (N·L≈0) barely → a FLOOR-selective purple-bright flash, temporally lockable to whatever gaps CurrCell/TryGetCell. Desk analysis found no per-frame gap trigger during plain corridor runs (UpdatePlayerCurrCell is stale-beats-null; the registry is stable in AP-36 dungeon mode) — but the branch inputs are exactly probed by [light] insideCell= lines, so one run settles it.
• T-B (flood with REAL production inputs): my harness feeds synthetic viewProj/eye. If the real per-frame inputs at the artifact moments differ (collided camera pressed into walls near seams, damped-forward gaze), the flood could still misbehave in configurations the sweeps missed. [pv-input] prints the exact inputs + flood count per frame.
• T-C: unknown-unknown (GPU state, driver, MSAA resolve…). If T-A/T-B both read clean at a flash moment → RenderDoc frame capture next.

Also found (real, filed, not these bugs)

• A8 double-sided shells stopgap still live: EnvCellRenderer.cs RenderModernMDIInternal maps CullMode.Landblock → CullMode.None ("while the architectural cause is isolated") — all cell shells draw two-sided. Perf + correctness debt; retire under the A7/A8 lighting arc.
• Lightning indoor gate missing (dormant): when weather strobes ship, SceneLightingUbo.Build needs the playerInsideCell gate or dungeons will strobe blue-violet.
• Flood follows floor-portals DOWNWARD from above (ScenarioC: under-hall network admitted at down-pitches). Retail's portal_side side test (0x005a59a0: portal_side≠0 → NEGATIVE side only; ==0 → POSITIVE only; IN_PLANE(±0.0002) → refused) appears to refuse this direction, but my plane-sign reading had unresolved contradictions (CCellPortal::UnPack normalizes portal_side at load, 0x0053ba1c/0x0053bc6a). OPEN QUESTION — harmless-looking (extra admitted cells draw below the opaque floor, z-buffer wins) but worth settling when the flood is next touched.

The live probe protocol (piggyback on the #175 door gate launch)

Env (add to the standard launch block):

ACDREAM_PROBE_LIGHT=1      # [light] insideCell/ambient/sun — rate-limited
ACDREAM_PROBE_PVINPUT=1    # [pv-input] one line/frame: exact flood inputs + count
ACDREAM_PROBE_CELL=1       # [cell-transit] timeline anchors

User reproduces in Facility Hub (~2 min): run the corridor across several seams until the purple flash shows; approach the stair room from the corridor, walk in, run down. Then close. Discrimination:

• Flash moments + [light] shows insideCell=False blips or ambient jumps → T-A confirmed (then root-cause the gap trigger from the same log).
• [pv-input] flood count drops (e.g. 12→1) at flash/pop moments while [light] stays clean → T-B confirmed (the line carries the exact inputs to replay in Issue176177FacilityHubFloodReplayTests).
• Stairs invisible while the flood contains 0x0182/0x0183 → draw-side hunt (RenderDoc next); absent → flood-side with the printed inputs.
• All clean → T-C: RenderDoc frame capture of a flash frame.

Apparatus shipped this session

• tests/AcDream.Core.Tests/Rendering/Issue176177DungeonSeamInspectionTests.cs — dat truth: portal polys/surfaces/draw verdicts, reciprocal coincidence, stair geometry owner, CellBSP containment, under-hall surface colors, DXT1 alpha histograms.
• tests/AcDream.App.Tests/Rendering/Issue176177FacilityHubFloodReplayTests.cs — production-matched flood replays: stair approach, descent, gaze sweep, corridor walk, and the ScenarioE incoherent-root sensitivity pin.
• tests/AcDream.Core.Tests/Physics/Issue176177SeamTransitLagTests.cs — resolver cell-flip position at the seam (plane-exact pin).