fix #176 #177: the camera-capped light snapshot evicted visible cells' torches — per-cell lighting popped at seams

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>

docs/architecture/retail-divergence-register.md

@@ -179,6 +179,7 @@ accepted-divergence entries (#96, #49, #50).
 | AP-80 | **PlanFromVelocity survives for velocity-only NPC cycles** (M16): UpdatePosition-derived speed picks Ready/Walk/Run cycles for server-controlled creatures whose UMs never arrive (scripted-path NPCs); retail derives every cycle from motion messages through the motion tables (R4-V4 note; pre-existing mechanism, row added per the V4 plan) | `src/AcDream.Core/Physics/ServerControlledLocomotion.cs` (`PlanFromVelocity`); consumer `GameWindow.ApplyServerControlledVelocityCycle` | Some ACE entities move by position updates alone — without this, they slide in T-pose; constants (StopSpeed 0.2, RunThreshold 1.25) tuned against live ACE traffic | Cycle-pick thresholds are acdream inventions — a creature intended to walk fast may show run legs near the threshold | retire in R6 (root motion + full per-tick order) |
 | AP-81 | **Remote-DR gravity toggled via the Gravity STATE bit**: the jump handler sets `Body.State \|= Gravity` at VectorUpdate and both landing blocks clear it after `HitGround()`; retail keeps GRAVITY set for the object's whole life and gates gravity ACCELERATION on the Contact transient (`calc_acceleration`) (pre-existing K-fix9/K-fix15 mechanism, row added during #161 — which also fixed the ordering so `Motion.HitGround()`'s verbatim `state&0x400` gate runs BEFORE the clear) | `src/AcDream.App/Rendering/GameWindow.cs` (VectorUpdate jump handler + the two landing blocks) | The DR tick integrates gravity only for airborne remotes; the flag dance delivers exactly that without porting the full contact-gated `calc_acceleration` chain; the #161 ordering fix keeps the retail HitGround contract satisfied | Any NEW call into `Motion.HitGround`/`LeaveGround` placed after the clear silently no-ops on the gravity gate (the #161 leg-2 class); grounded remotes carry a non-retail state word (probes comparing state bits vs retail mislead) | `CPhysicsObj::calc_acceleration` (contact-gated); `set_on_walkable` 0x00511310; retire in R6 (contact-gated accel + persistent GRAVITY) |
 | AP-82 | **StickyManager deep-overlap back-off sign pin**: when the stick-gap overlap exceeds one tick's step (`speed×quantum < \|dist\|`, `dist < 0`), acdream applies `delta = −(speed×quantum)` (rate-limited back-off); ACE's literal port keeps `+delta` there — a runaway that steers INTO the target with equilibrium at centers-coincident. The BN mush (0x00555554-0x00555597) is unreadable on exactly this compare; the pin is refuted-by-evidence against ACE-literal: #171 gate-3 probe showed 1661 deep-overlap ticks all steering inward (monsters converged to centerDist≈0 — "monster inside the player") while retail side-by-side on the same ACE shows separation. ACE servers essentially never reach the branch (quantum ≥1/30 → threshold ~1 m; render-rate quanta → ~0.13 m) | `src/AcDream.Core/Physics/Motion/StickyManager.cs` (`AdjustOffset` delta clamp; conformance `StickyManagerTests.AdjustOffset_DeepOverlap_BacksOff_RateLimited`) | Minimal interpretation consistent with the mush structure AND observed retail; identical to ACE-literal in every shallow/outside case | If retail's true deep-overlap behavior differs (e.g. no movement at all), our back-off rate diverges in that rare state; verify via cdb `StickyManager::adjust_offset` trace with a forced overlap when convenient | `StickyManager::adjust_offset` 0x00555430 (x87 mush); ACE StickyManager.cs:117-121 (the literal branch this pin overrides) |
+| AP-85 | **Per-frame flat point-light snapshot with a 1024 nearest-to-camera safety cap** (#176/#177 fix, 2026-07-06): acdream collects ALL lit point/spot lights into one flat per-frame snapshot (`BuildPointLightSnapshot`) that per-cell/per-object selection (`SelectForObject`, the faithful AP-16 8-cap) draws from; retail registers lights per-CELL (`insert_light` 0x0054d1b0) and consults the reaching set with NO global pool cap. The previous cap of 128 BIT in the Facility Hub (366 registered fixtures → 238 camera-distance evictions/frame → in-range torches of VISIBLE cells dropped from their 8-sets → per-cell Gouraud lighting popped with camera movement = the #176 purple seam flash + the #177 stair-room light pop-in) | `src/AcDream.Core/Lighting/LightManager.cs` (`MaxGlobalLights`, `BuildPointLightSnapshot`); pin `LightManagerTests.PointSnapshot_HubScaleLightCount_ObjectSelectionIsCameraInvariant` | The flat scan is a perf-shape adaptation (feel-identical while the cap never bites); per-object selection stays retail's `minimize_object_lighting` 8-nearest | If content ever registers >1024 lit lights in relevance range, per-cell sets silently become camera-dependent again — the #176/#177 pop class returns. Raise the cap, don't tune it | `minimize_object_lighting` 0x0054d480 (no global pool cap); `insert_light` 0x0054d1b0 (per-cell registration) |
 | AP-84 | **BSP shadow-shape part poses = motion-table default-state frame snapshot at registration, not retail's live CPhysicsPart pose** (#175): server entities with a wire MotionTableId register their BSP part shapes at the default style's first-cycle LowFrame pose (the closed pose for doors — `GameWindow.MotionTableDefaultPose`); retail collision reads each part's CURRENT pose every test. Equivalent for the door lifecycle (closed = default pose; open = ETHEREAL bypasses collision entirely, #150) and for idle statics | `src/AcDream.App/Rendering/GameWindow.cs` (`MotionTableDefaultPose` + the RegisterServerEntityCollision override); `src/AcDream.Core/Physics/ShadowShapeBuilder.cs` (`partPoseOverride`) | Registration is one-shot in acdream (retail re-poses parts per frame); the default-state pose is the correct idle pose and the only non-ethereal pose doors ever collide in | An entity whose server-driven motion state materially MOVES a BSP-bearing part while NON-ethereal would collide at the stale default pose (no known case — doors are the dominant BSP-part weenies); revisit if animated non-ethereal BSP movers appear | `CPhysicsPart` live pose (see #150 notes); motion-table default state = CPartArray init; ShadowShapeBuilder placement-frame fallback for table-less entities |
 | AP-83 | **CylCollideWithPoint PerfectClip TOI sub-branches decoded via ACE, not the binary**: the CCylSphere family port (2026-07-05, retires AP-6) reads `collide_with_point`'s PerfectClip time-of-impact math (0x0053adb6+) from ACE `CylSphere.CollideWithPoint` because the BN x87 mush is unreadable there; two ACE-verbatim quirks ported as-is (`movement.Z + radius` in the not-definite ascending case; `GlobalCurrCenter[0]` used even for head-sphere hits — the latter matches the raw decomp read). NOT exercised in M1.5: no mover sets PerfectClip (players never do; the non-PerfectClip path — SetCollisionNormal + Collided — is decomp-verified). Separately, the grounded head-sphere slide passes the HEAD disp per retail 0x0053b843 where ACE passes the foot disp — retail wins (ACE bug, not copied) | `src/AcDream.Core/Physics/TransitionTypes.cs` (`CylCollideWithPoint`; pseudocode doc `docs/research/2026-07-05-ccylsphere-collision-family-pseudocode.md` §7-8) | The load-bearing paths (non-PerfectClip Collided; the family's step-up/step-down/land) are decomp-verified; the TOI tail is dead code until missiles arm PerfectClip | If missiles (F.3) arm PerfectClip, the two ACE quirks may diverge from retail — clip-through or wrong deflection on cylinder targets; re-decompile 0x0053acb0 in Ghidra before shipping missiles | `CCylSphere::collide_with_point` 0x0053acb0 (pc:324173, x87 mush from 0x0053adb6); ACE CylSphere.cs `CollideWithPoint` |