Release Healer Man 0.1.26

PLAYTEST_NOTES.md

@@ -1854,3 +1854,141 @@ Vite's existing large-chunk advisory remains informational.
 - Verification passed all 1,133 Vitest cases across 208 files. The production
   build completed with 974 modules transformed; arena, KTX2, and environment-
   performance audits remained green, and the release-candidate check passed.
+
+# 2026-09-09 — dungeon wall and nameplate performance regressions
+
+- The 0.1.24 material fallback honored BLEND tags on legacy WoW WMO walls,
+  floors, and ceilings, disabling their depth writes. WoW environments now use
+  their authored material roles again; RuneWaker environments retain their
+  source MASK/BLEND semantics. Shipped Wailing Caverns wall materials, vines,
+  and water are covered by regression tests.
+- The 0.1.25 boss-name occlusion scanned the complete rendered scene with
+  Three.js raycasts. A live Kalin's Shrine CPU profile was dominated by triangle
+  intersection and vertex decoding. Boss names and nearby NPC prompts now use
+  bounded camera-to-label queries against Rapier's existing physical spatial
+  index, at most ten times per second, skipping offscreen labels. Walls, floors,
+  and ceilings occlude labels; actors and trigger sensors do not.
+- In an isolated offline Edge playtest at 1440 x 900 on an AMD RX 9070 XT,
+  Kalin's Shrine improved from about 10 FPS (98.5 ms/frame, 167 draw calls,
+  DPR 1.125) to 55.5-56.1 FPS (17.8-18.0 ms/frame, 168-170 draw calls,
+  DPR 1.5). This is a representative entrance/combat comparison, not an audit
+  of all thirty RoM dungeons. The environment alone took about 0.8 ms CPU and
+  0.5 ms GPU with 14 draw calls.
+- Wailing Caverns screenshots show continuous opaque cavern walls and ceilings.
+  A controlled toggle of the old material flags reproduced the missing wall
+  sections at the same camera position; restoring depth-writing materials fixed
+  them. Its full scene ran at about 57 FPS. No uncaught client errors appeared.
+- Evidence and repeatable local benchmark helpers are under
+  `playtest-artifacts/2026-09-09-dungeon-rendering/`: `kalin-before.cpuprofile`,
+  `kalin-after.cpuprofile`, `kalin-after-metrics.json`, `kalin-before.png`,
+  `kalin-after.png`, `wailing-regression.png`, and `wailing-restored.png`.
+- All 26 focused material, label occlusion, combat sight, aggro visibility,
+  instancing, and NPC-prompt tests passed. Physical-occlusion tests exercise
+  actual Rapier worlds rather than just source-code assertions.
+
+# 2026-09-09 — campaign rendering audit and Thor performance preparation
+
+- Audited the actual visual GLB headers for all 54 WoW campaign dungeons:
+  9,402 materials, including 3,586 legacy BLEND materials restored to opaque
+  rendering by the shared policy. Every WoW package contained affected
+  materials. A permanent shipped-asset test covers all 54 definitions and
+  retains explicit cutout, blend, and additive roles. This is material-policy
+  coverage, not a claim that every room has been visually traversed.
+- Live screenshots checked Wailing Caverns, Shadowfang Keep, Utgarde Pinnacle,
+  the Nexus, and Tinkertech Showdown. The sampled walls, floors, stairs, and
+  ceilings remain solid; authored lighting effects remain transparent. Kalin's
+  Shrine and Forsaken Abbey provided RoM rendering checks. Roster and in-dungeon
+  screenshots show character bodies and equipped weapon silhouettes after the
+  shared character animation change.
+- A controlled Utgarde Pinnacle entrance test found about 1,739 draw calls,
+  1.95 million triangles, and 30.5 FPS on the desktop GPU. Clipping at the
+  existing full-fog distance reduced this to about 209 calls, 567,000 triangles,
+  and 60 FPS at the same viewpoint. The camera now follows each dungeon's fog
+  distance, including transitions that retain the existing Canvas/camera.
+  The background uses the fog color to avoid a clipping seam at the horizon.
+  All 84 dungeon definitions have render-distance coverage.
+- Static dungeon visual transforms are cached until the root or its parent's
+  world transform changes. Tests verify imported matrices, repeated unchanged
+  frames, root movement, and parent movement. Characters, moving world objects,
+  and physics remain outside this immutable visual cache.
+  The final 63-second Utgarde sample held 60 FPS with 213 draw calls:
+  main-thread work averaged 9.2 ms/frame, renderer CPU averaged 7.5 ms
+  (p95 8.3 ms), and GPU averaged 2.7 ms (p95 4.1 ms). Movement and jump input
+  continued to animate the player while the environment remained stationary.
+- Character locomotion and jump clip selections are cached by animation
+  library and weapon stance. The final Kalin CPU profile contains no sampled
+  per-frame base-clip selection/name parsing. Frame-cap deadlines now tolerate
+  sub-millisecond timestamp rounding without drifting or emitting catch-up
+  frames; tests cover 60 FPS on rounded 60 Hz and 120 Hz callbacks.
+- A 31.5-second Kalin sample with a connected browser companion display held
+  60 FPS: main-thread work averaged 8.7 ms/frame, renderer CPU averaged 5.4 ms
+  (p95 5.9 ms), and GPU averaged 2.4 ms (p95 3.4 ms). The companion used about
+  5.1 ms of main-thread time per second. Recent ability-cost snapshot work was
+  not a dominant hotspot in this sample. Startup warmup is bounded to loading;
+  companion snapshots are already limited to ten per second during gameplay.
+- These are isolated offline Edge desktop samples on an AMD RX 9070 XT at
+  1920 x 1080 CSS pixels, DPR 1.5 (2880 x 1620 renderbuffer). Party composition
+  can change between dungeon entries. They do not certify Android performance
+  or sustained busy-combat headroom. ADB listed no connected Android device.
+- Thor acceptance should use the actual model, both native displays, target
+  resolution, and a sustained 20-minute traversal/combat run after warmup.
+  Provisional headroom target: p95 main-thread and GPU work each at or below
+  12 ms, with p99 presented frame intervals within the 16.7 ms 60-FPS budget.
+  Test large pulls, boss effects, menus, and repeated dungeon transitions;
+  record thermal behavior and frame-time spikes, not only average FPS.
+  The environment asset audit still has 60 grandfathered budget exceptions;
+  its green result means no new asset regression, not a Thor FPS guarantee.
+- Evidence is in `playtest-artifacts/2026-09-09-thor-audit/`, including roster,
+  dungeon, and companion screenshots, CPU profiles, and CPU/GPU JSON samples.
+  The full material report is
+  `playtest-artifacts/2026-09-09-dungeon-rendering/wow-material-audit.json`.
+- Validation passed 227 focused tests covering avatar rendering and animation,
+  camera distances, material policies, frame pacing, static transforms, label
+  visibility, and companion snapshots. The final static-cache change also
+  passed the 105-test scene subset. The production build and release-candidate
+  check passed; the existing large-bundle advisory remains.
+
+# 2026-09-09 — additional CPU headroom from culling and shared skeletons
+
+- Prepared dungeon visuals now group static siblings into conservative spatial
+  bounds. One scene hook checks these bounds after the current camera and world
+  matrices update, before Three.js traverses individual meshes. Bounds include
+  complete wall geometry and all instances; lights, animated meshes, and meshes
+  explicitly disabling frustum culling remain outside these groups. The hook
+  supports streamed chunks and restores its predecessor after cleanup.
+- At the same Utgarde Pinnacle entrance, only 16 of 1,383 spatial groups were
+  visible. The renderer examined 108 static meshes instead of 7,002, retaining
+  the same 206 draw calls and 566,913 rendered triangles. Controlled culling
+  toggles produced byte-identical canvas PNGs across eight camera headings.
+- In the same running scene, spatial grouping reduced average main-thread work
+  from 9.45 to 6.83 ms/frame, and renderer CPU from 7.81 to 5.31 ms/frame.
+  Removing containers emptied by static instancing reduced the final fresh-run
+  averages further to 6.49 ms main-thread and 4.85 ms renderer CPU (p95 5.2 ms).
+  The final sample lasted 87.8 seconds and held 60 FPS. Authored empty scene
+  nodes are preserved; only containers emptied by our batching are removed.
+- The shared character renderer still uses SkeletonUtils.clone, then reuses
+  equivalent skeleton palettes within that owned clone. Bone identities, order,
+  and inverse bind matrices must all match; per-mesh bind transforms remain
+  unchanged. Different characters retain independent bones and animations.
+  Owned bone textures are now explicitly disposed during character teardown.
+- A live character audit found 66 skeleton objects for one player palette and
+  256 objects for four party palettes. A Wailing Caverns toggle removed 309
+  duplicate skeletons and produced an identical rendered image. Tests verify
+  posed vertices, differing bind transforms, incompatible palettes, independent
+  cloned characters, and redundant texture disposal.
+  In the same Wailing scene, main-thread work fell from 7.40 to 6.70 ms/frame;
+  renderer CPU fell from 5.51 to 4.86 ms/frame, and allocated GPU textures fell
+  from 292 to 217. This measures palette sharing after spatial culling was
+  already enabled. Both samples held 60 FPS.
+- The final built runtime showed a complete roster body and staff with one
+  skeleton for 66 skinned meshes. Kalin's Shrine showed one player palette and
+  four independent party palettes; walking retained the player body and weapon
+  silhouette. Wailing Caverns and Utgarde screenshots retained solid dungeon
+  surfaces. There were no uncaught client errors during these playtests.
+- All 210 targeted avatar, visibility, instancing, material, camera-distance,
+  and label tests passed. The production build and release-candidate check
+  passed. The existing large-bundle advisory remains.
+- Evidence is under `playtest-artifacts/2026-09-09-static-culling/`: paired PNGs,
+  eight-heading pixel comparisons, profiles, before/after timing JSON, and
+  screenshots. Measurements use the same desktop Edge/RX 9070 XT setup as the
+  preceding audit. These CPU improvements do not certify Thor performance.