Update Healer Man server

PLAYTEST_NOTES.md

@@ -4,10 +4,291 @@ This file is the durable record of runtime behavior, regressions, and the
 evidence required before calling a visual bug fixed. Update it whenever scene,
 camera, input, physics, mob-rendering, or asset-loading code changes.
 
-New routine live tests use the full **Stormwind Stockade** (`stormwind-stockade`)
-under [the current test policy](docs/TEST_DUNGEON_POLICY.md) and shared defaults in
-`scripts/playtest-defaults.cjs`. Dated Wailing Caverns captures below remain
-historical evidence and do not select future tests.
+New routine live tests use **Stormwind Stockade: Cellblock Rush**
+(`bite-sized-stormwind-stockade`) under [the current test policy](docs/TEST_DUNGEON_POLICY.md)
+and shared defaults in `scripts/playtest-defaults.cjs`. Each run selects one
+Stockade boss from its shared run ID; record the boss actually observed. Full
+Stockade campaigns below remain dated evidence and require explicit selection
+for future runs. Dated Wailing Caverns captures remain historical evidence.
+
+## NPC utility/control completion — 2026-10-06
+
+Guarded ordinary-Creature Critical Block self-imports, exact WoW335 Shield
+Bash 72 player/party utility blocking and source shout speed/fear immunity are
+implemented. No real Critical Block NPC assignment is invented. Shield Bash's
+trigger 29703 remains in the parent descriptor, but its Daze child is absent
+from the runtime catalog. General shout Magic hit/resist/deflect and exact fear
+locomotion remain gaps. Source audit leaves Ascension stance DUMMY +42 and
+Targorr Thrash timing gated. See
+[the completion report](docs/ENEMY_UTILITY_CONTROL_COMPLETION_2026-10-06.md).
+
+One final routine run uses the frozen build and **Stormwind Stockade: Cellblock
+Rush**, `playtest-artifacts/stockade-weapon-utilities/20261006-141251022/`.
+The normal UI-created Discipline Priest profile receives isolated pre-entry
+level 30 preparation and four Ready companions. The actual seeded boss is
+Targorr 1696, source/live `creature-84027`; native progress/results agree on 1/1
+in 03:09.2, no deaths and 2,337 effective player healing. This is the one-boss
+slice. Committed evidence retains 749 consecutive events (IDs 9–757), with zero
+gaps/errors/capacity loss: 56 source mainhand, four offhand, two Thrash
+extra-mainhand outcomes and 172 player/party hand results. Three Disarm windows
+contain seven actual hand results. Block/shout/Bloodlust auras are unobserved;
+focused fixtures provide their bounded coverage.
+
+Root visually checks roster, equipped silhouette, mid-route and results:
+character bodies and staff/shield/blade silhouettes are present. Overlap/HUD
+occlusion limits complete per-companion gear review. There are no page
+exceptions; only favicon 404. The one helper/server/browser is cleaned up by
+the harness. No production source changes follow the frozen live run.
+
+Build SHA-256
+`8cf9513613aa3ef6a0a64de392e90521b0349d85311f930ba8808e868d6867d9`
+covers 5,967 files. Reused isolated content manifest SHA-256
+`ac0fa2a867ba0a495c172de2b89532d69b32d56d032482b26020ce2d189f074b`
+remains unchanged; 2,124 served manifest-referenced objects verify with zero
+integrity failures. Raw response bodies are not retained. Ninety-six focused
+runtime/helper tests across 15 files, six source-pin Node tests, Ready-five
+(7 Vitest/20 Python), avatar 56, final TypeScript and production build pass.
+No loot wiring, level-cap or readiness change is made. Subprocesses use only
+GPT-6 Luna extra-high; root completes the utility runtime tests after Luna
+capacity prevents that follow-up from running.
+
+The evidence-audit follow-up also encounters Luna capacity. Root's clearly
+labeled `root-evidence-audit.cjs/.json` rehashes all 5,967 build files, seven
+loaded chunks and 2,124 served manifest objects (77,019,550 bytes), and
+recomputes event-ID continuity and native completion. It passes with zero
+integrity failures. This is root verification, not a completed independent
+subprocess audit.
+
+Final release check passes 478 changed paths including untracked files; game
+and vault diffs are checked separately and all 13 task documents pass EOF/
+whitespace checks. The task worklog is closed.
+
+## 2026-10-05 CoA Brewing companion and Splash integration
+
+- Brewing healer companions now prepare learned Frog Bones during healing
+  downtime through normal AI selection/payment. An existing brew, injured party,
+  insufficient Mana, cooldown, resurrection/dispel demand or enemy cast blocks
+  preparation. Preparation uses an authored 90% health threshold and does not
+  claim an optimal live-server rotation. Ordinary healing scoring uses the
+  caster's projected ingredient state and excludes ingredient casts.
+- Splash Potion now uses friendly ground targeting and a shared destination for
+  its heal and sourced Frog Bones absorb. The importer no longer grants inactive
+  ingredients or unresolved SLS combos unconditionally. Actual player/NPC tests
+  cover range, nearby/distant recipients and naturally prepared Potion Toss.
+- Guardian formation actions are absent from the promoted runtime catalog;
+  Footman's Calling cannot use an assumed active formation. Paragon's August
+  builder and July DBC payload conflict. Exact next integration boundaries and
+  remaining Brewing rider evidence are in `docs/COA_FIRST_PARTY_MILESTONE.md`.
+- 37 focused tests and Ready-five checks pass. The five chosen level-60/seed-1
+  Stockade/Hamhock seat diagnostics record no deaths; Barbarian defeats at 13.6s
+  and Pyromancer at 24.8s, other seats reach 30s. The Guardian-seat Brewing NPC
+  records one naturally selected Frog Bones cast. Low pressure limits healing
+  and tank conclusions. Raw evidence is under
+  `artifacts/coa-first-party/stockade-integration-2026-10-05/`. All five remain
+  Needs Testing; no live browser playtest was performed in this integration.
+- Final production build and release checks pass, including source/runtime/UI
+  targeting parity and preservation of unrelated CoA records. The canonical
+  CoA vault note records the new owners and source boundaries.
+
+## 2026-10-05 CoA first-party mechanics and rotation follow-up
+
+- The selected Guardian/Vanguard, Witch Doctor/Brewing, Barbarian/Brutality,
+  Ranger/Archery and Pyromancer/Incineration remain Needs Testing. Source evidence
+  comes from the cached August Vol'Jin capture and global DBC; no new crawling or
+  downloads were needed. See `docs/COA_FIRST_PARTY_MILESTONE.md` for ownership,
+  source IDs, exact limits and the next mechanics work.
+- Guardian Vanguard now gains actor-owned stacks on normal swings and blocks,
+  caps at ten, refreshes for 60 seconds and reduces each magic school by 2% per
+  stack. Player and companion tests use real incoming blocks and Hamhock's
+  imported Chain Lightning: two stacks reduce the hit by 4% before one stack is
+  consumed. Ordinary melee abilities do not add stacks. Periodic consumption
+  remains unspecified and excluded.
+- Brewing stores actor-owned ingredients, gates one/two capacity before payment
+  and adds Frog Bones' sourced eight-second shield to Potion Toss's selected
+  ally. Actual player and companion cast paths are checked, including removal
+  of the bogus caster shield, foreign/absent marker rejection and unchanged
+  resources/cooldowns on a full-cap rejection. The companion Toss test primes
+  its ingredient; default healthy AI ingredient preparation remains unfinished.
+  Other ingredient riders and unspecified replacement/removal remain gaps.
+- Learned Beheader restores maximum Energy after an effective Decapitate kill
+  for a player or companion. Nonlethal, rejected and unrelated kills do not
+  refund. Additional-Energy conversion and generic weapon-damage projection
+  remain separate work.
+- Ignite now uses its sourced unit target rather than spreading its burn via
+  effect-radius metadata. Unit periodic maintenance compares ability, caster,
+  recipient and remaining time; another target's effect does not block the new
+  cast. Mojo Beam's incidental caster aura no longer redirects the friendly
+  unit-heal decision to the caster. No optimal priority list or clipping window
+  was inferred. An actual companion Mojo Beam test checks authoritative Mana
+  payment and its first pulse healing the injured Vanguard ally, not the caster;
+  the full duration conflict remains unresolved.
+- Earlier three-cast Ranger/Pyromancer companion counts were a reporting defect,
+  not observed inactivity. The collector now counts party releases and dedupes
+  channels. Final level-60 seed-1 Hamhock diagnostics run only this explicit
+  party, with every active-player seat and a 30-second limit. All finish without
+  deaths; Barbarian defeats at 13.6 seconds and Pyromancer at 27.8, while the
+  other seats reach the limit. Healer-seat companion counts are Ranger 19 and
+  Pyromancer 16; Ranger still records 32.3% Focus starvation. Outcomes and low
+  pressure do not establish tank/healer coverage or balance. Raw ignored runs
+  are `artifacts/coa-first-party/stockade-2026-10-05/seat-1-after.json` through
+  `seat-5-after.json`; two/four also retain before runs.
+- This is automated mechanics/diagnostic evidence. No live browser playtest,
+  readiness promotion, full repository suite or broader class sweep occurred.
+- Focused checks pass 126 tests in 19 directly affected files, plus the existing
+  seven Ready-five Vitest and twenty Python checks. Both generated Ignite actions
+  are unit-targeted; the other twenty class records, resources and triggered
+  payloads are preserved.
+- Final production build passes TypeScript, asset audits, bundling, compression
+  and bundle verification. Final release hygiene includes untracked files and
+  passes; the canonical CoA vault note is updated in its separate repository.
+
+## 2026-10-05 Enemy stances and NPC blocking
+
+- NPC Defensive/Battle Stance now use the pinned form and linked passive data,
+  replace each other, and clean up on expiry/reset/evade. Defensive multiplies
+  all-school damage taken by 0.9 and dealt by 0.95. Improved Blocking adds 75
+  percentage points for six seconds. Reviewed Stockade Inmates prepare Battle
+  Stance after one idle second; Kam's combat self auras retain source timing.
+- Player and companion melee/ranged autos share ordinary source-backed WoW
+  creature blocking: base chance 5%, block value floor(level / 2), one skill
+  adjustment, actual facing and source NO_BLOCK/cast/control/evade gates. Spell
+  and periodic blocking attributes, hostile outgoing threat and Ascension's
+  unsupported stance DUMMY rider remain separate gaps. Native RoM defenders
+  retain their existing model. The extra incapacitation labels are documented
+  simulator approximations; plain root still permits blocking.
+- All 102 affected Vitest checks, seven Ready-five Vitest/twenty Python checks,
+  seven importer cases and the production build pass. The source comparison
+  preserves unrelated encounter health, placement and attacks. No selected
+  imported template sets NO_BLOCK; its evidence is the importer/helper fixture.
+  See `docs/ENEMY_STANCES_BLOCKING_2026-10-05.md` and the scoped live report.
+- Normal full Stockade run `20261005-184333429` confirms all five exact boss
+  IDs and the native full-clear screen, Easy/Normal, at 24:25.4. The level-30
+  Discipline Priest/four Ready companions finish alive without observed deaths
+  or resurrection recoveries. Natural progression ends at player 40 / party
+  39. Frozen SHA-256 remains
+  `1fcd021af0ef90c8d994cf1ab853ef0e62f4d4767fa3acaab698b14121194310`.
+- Live aura state contains both NPC stances and Kam's Improved Blocking.
+  There are eight party-to-hostile and 141 hostile-to-party block outcomes.
+  Only 33 outcomes have a sampled latest-result/raw-amount match for blocked
+  amount; unmatched values remain unknown. Melee/player/native-RoM attacker
+  cases and exact timing boundaries use deterministic integration evidence.
+- Root visually reviewed the 25/50/75-percent captures and full-clear screen:
+  bodies and equipment silhouettes render, with existing overlap/edge occlusion
+  still present. No visual or race-model fix is claimed. Page errors are zero;
+  the console has one missing-favicon 404. All 2,124 fetched hash-addressed
+  objects verify, but the broad source content manifest changed during QA;
+  whole-content immutability is not claimed. The scoped report records that
+  limit and retains the superseded observability-limited first attempt.
+
+## 2026-10-05 Stockade enemy mechanics and Chronomancer
+
+- Enemy Chain Lightning 421 now selects three distinct living visible targets
+  with nearest-target jumps, profile-specific decay and geometry independent
+  from home-leash aggro. Source Rend 6547/16509 applies five three-second bleed
+  pulses, restarts its ordinary refresh phase and snapshots caster damage-done
+  modifiers while keeping recipient damage-taken modifiers dynamic.
+- Verified Enrage and shouts now use shared auras, source duration/radius,
+  physical damage or attack-power channels, source stacking and melee swing
+  scheduling. Stance forms and Improved Blocking were unsupported at this
+  stage; the later stance/blocking continuation above resolves their ordinary
+  NPC paths. Chrono Sorcery's structured +12% Mana
+  restores the existing exact 414 Mana expectation; source power identity keeps
+  Rage/Runic Power riders separate from Mana.
+- Required checks pass: 135 affected Vitest, two selected original combat
+  regressions, seven Ready-five Vitest and twenty Ready-five Python checks.
+  The production build passes all hooks and the final standalone TypeScript
+  check passes. Final release hygiene and game/vault diff checks pass. No full
+  suite or expanded balance sweep was run.
+- Live QA uses the immutable `frozen-dist-20261005-153220` copy: all 5,964 files
+  aggregate to SHA-256
+  `296f26c1a58d096b925977f8a0f4aebc5a2de20ba4ce804592188a87102f99e0`.
+  It contains `index-D-xYtSuF.js` and `combatStore-BmIrp-s1.js`. The server's
+  content source remains the read-only project catalog; downloaded objects and
+  account data use a disposable browser/server profile.
+- Final normal full Stockade run `20261005-160913676` verifies all five boss IDs
+  and the native full-clear screen at 21:32.5, Easy/Normal. All five party actors
+  finish alive; the tank has three observed casualties followed by normal
+  resurrection UI recoveries, with no observed player death. The results
+  overlay ends route input at checkpoint 462/481. This is not balance
+  certification.
+- Live evidence contains four unique Hamhock Chain Lightning cast windows,
+  including a three-recipient Nature damage group of 579/347/208 raw damage;
+  359 Rend damage events across both spell IDs; Targorr's source Enrage and
+  Battle Shout applications to hostile allies. Demoralizing Shout has 44 unique
+  sampled cast windows; its recipient effects and the unobserved
+  source-profile/refresh/expiry boundaries use deterministic proof.
+  Source-aura coverage includes shared combat-aura telemetry even when the
+  legacy `mob.auras` field is empty.
+- Root visually inspects equipped-entry, 50%, 75% and final-results captures
+  from `campaign/final-build-20261005-160915922/` under that run. Bodies, HUD and
+  full-clear results render. Actor overlap and partial occlusion persist; no
+  visual fix is claimed. Frozen-build/content-manifest fingerprints are stable
+  and all 2,124 served content objects match their hash addresses. Page errors
+  are empty; the console contains only the optional favicon 404.
+- Preserve the superseded-before-entry, content-source 404, one-boss movement
+  overshoot and serializer-failure attempts. Their harness issues are corrected
+  in the final wrapper; the scoped report retains their original evidence.
+- See [the coordinated report](docs/STOCKADE_ENEMY_MECHANICS_2026-10-05.md)
+  and [the scoped live report](docs/STOCKADE_ENEMY_MECHANICS_PLAYTEST_2026-10-05.md).
+
+## 2026-10-05 RoM healing, party complements and player Auto Shot
+
+- Group Heal's party-heal effect now carries its sourced 12 m radius for both
+  player and NPC resolution. Generated native-five parties exclude the active
+  player's ordered identity when the other four can fill the requested roles.
+  Scout/Rogue recurring Auto Shot uses 80% ranged weapon DPS, native or shared
+  weapon cadence, owner Blood Arrow costs and ordinary start/stop bindings.
+  It uses the existing resolved Training Gear bow with empty inventory and
+  supports compatible custom-item overrides.
+- Required checks pass: 96 affected Vitest plus 7 Ready-five Vitest and 20
+  Ready-five Python checks. The production build and its pre/post hooks pass.
+  Supplemental generic combat checks pass 110/111; the Chronomancer resource
+  mismatch (expected 414, actual 369) also reproduces with committed HEAD
+  versions of the changed production modules. See the source-only replay in
+  `artifacts/rom-healing-ranged-baseline/README.md`. The enemy-mechanics follow-up
+  above corrects that resource mismatch without changing the original assertion.
+- Bounded full Stockade run `20261005-042203897` passes the three required
+  live criteria in 61.051 seconds on `index-Cvmh-YKE.js`. Its recursive 123-JS
+  manifest remains unchanged throughout:
+  `ed3cb65752426810c84be15184306114fdf36d3804a23f917213fc854f4651d4`.
+  Both normally generated rosters are the exact native-stat complement of the
+  Priest/Knight and Scout/Rogue players. Profiles are disposable and their
+  level 30/secondary class are prepared only before entry; combat uses normal
+  UI input with read-only telemetry.
+- Normal Attack party order produces natural injuries. A player Group Heal
+  starts/completes and heals Caelan for 92 effective HP; its full-health self
+  event has zero effective healing. NPC recipient position is unavailable in
+  that capture, so it does not establish the 12 m boundary live. Deterministic
+  player/NPC integration owns distance, living and LOS proof.
+- Normal Auto Shot input produces two distinct Physical enemy hits of 37
+  effective damage, 2,612 ms apart. Stop input clears the mode and its observed
+  window has no player Auto Shot hit. The captured restart is inconclusive as
+  the target dies. Owned Blood Arrow is active; aggregate HP samples overlap
+  combat/recovery and do not prove exact self-cost. Focused integration proves
+  the once-per-attempt 2% max-HP cost with no MP or separate self-damage event.
+- Root inspected `priest-knight-group-heal-after-observation.png`,
+  `scout-rogue-auto-shot-target-1.png` and
+  `scout-rogue-auto-shot-cancel-restart.png` from that run. Bodies, equipped
+  silhouettes and HUD render; doorway/party overlap and camera occlusion
+  persist. No page or asset HTTP errors occur; the console has a favicon 404.
+  This is a bounded mechanic diagnostic, not a new full-clear or balance proof.
+- Preserve initial run `20261005-040846797` (fixture name casing) and
+  `20261005-040932648` (inconclusive combat). The latter exposed the corrected
+  explicit-assignment-only guard and selection-only Defend encounter setup.
+  Earlier Stockade captures retain their original build identities. Current
+  Blood Arrow cap-80 scaling remains unverified at the retained 10% fallback;
+  Auto Shot's unquantified last-offensive-skill bonus remains omitted. See
+  [the coordinated report](docs/ROM_HEALING_PARTY_RANGED_FOLLOWUP_2026-10-04.md)
+  and [the live report](docs/ROM_HEALING_RANGED_STOCKADE_PLAYTEST_2026-10-04.md).
+- Supplemental same-build run `20261005-042845955` passes all required criteria
+  again in 61.658 seconds, with 92 effective Group Heal HP and two Physical
+  Auto Shot hits of 36 each. Ordinary Stop party order precedes cancellation;
+  no shot occurs in the observed stop window, but target death again prevents
+  restart. Deterministic integration verifies stop-window silence/no cost and
+  a fresh full-interval deadline after reactivation without an immediate or
+  catch-up hit. The original Defend order is restored. Root also inspects that
+  run's `scout-rogue-auto-shot-cancel-party-stopped.png` and
+  `scout-rogue-auto-shot-cancel-restart.png`; the same overlap/occlusion remains.
 
 ## 2026-10-04 Stockade default and Blood Arrow toggle
 
@@ -21,7 +302,8 @@ historical evidence and do not select future tests.
   right-click removal starts a ten-second cooldown; unrelated buff clicks leave
   Blood Arrow active. Permanent NPC Blood Arrow icons remain visible without
   exposing every long-term buff. Exact costs and rider lifetime have focused
-  integration coverage; player recurring ranged Auto Shot remains open.
+  integration coverage; player recurring ranged Auto Shot was still open at
+  this captured build and is implemented in the 2026-10-05 follow-up above.
 - Corrected-build ordinary Stockade run `20261004-194732821` passed both owners'
   cancellation, cooldown rejection/reactivation/reuse and unrelated-buff guards,
   plus owner-only NPC cancellation. Scout/Rogue's isolated Shot consumed 46 HP
@@ -31,7 +313,7 @@ historical evidence and do not select future tests.
   The image named `knight-scout-body-and-gear.png` shows the training-gear panel.
 - The primary Knight/Scout roster has all five active identities. The optional
   Scout/Rogue damage-role roster instead duplicates Scout/Rogue and omits
-  Mage/Warlock; companion player-pair exclusion is tracked in the backlog.
+  Mage/Warlock; that captured selection gap is fixed in the 2026-10-05 follow-up.
 - The first full route exposed unsupported NPC utility spells dealing damage.
   Two source corrections now omit thirteen verified utility/extra-attack IDs.
   The second pass removes 51 more false damage rows and corrects 30 real
@@ -4668,3 +4950,621 @@ The equipment/manual-formula, damage-pressure and production QA evidence above s
 - Visually reviewed the fresh-profile `playtest-artifacts/party-equipment/basic/RoM-roster-basic.png` and `RoM-Caelan-basic.png`: the roster has the original textured wooden staff and clothing; Caelan holds the original bow upright with the string side toward him. The bow-facing dot product is 0.896 and its vertical-axis alignment is 0.995. The entire fresh-profile route passed with zero page, console, or non-favicon HTTP errors; no optional native character pack was installed. Results are in that directory's `results.json`.
 - The completed native-pack WoW/RoM route also reported zero browser or asset errors. Reviewed `playtest-artifacts/party-equipment/WoW-party-front.png`, `RoM-roster.png`, and the native close-up follow-up `RoM-Caelan-native.png`/`RoM-party-front.png`: WoW robes, shoulders, armor and cloaks remain visible; the native RoM roster shows the complete source staff and outfit; Caelan carries the original textured bow in the native lowered idle grip, with the limb tips/string side toward his torso. The final native-facing dot product is 0.465. The close-up uses the normal camera's collision-safe position and a temporary narrower field of view; a previous actor-relative camera position landed inside a wall and was rejected as visual evidence. Nearby native companions show the original shield, hammer and daggers. The group remains crowded on the entry stairs, so the captures do not certify party spacing or every combat animation.
 - Focused validation: all 56 avatar tests pass, including missing-pack equipment resolution, source bow alignment, and inverse socket-frame coverage. The final production build, asset audits, compression, and release-bundle verification pass. No original assets needed regeneration, and no deployment was performed.
+
+## 2026-10-05 — RoM floor/wall texture and lighting pass
+
+- Reviewed all 30 registered base RoM environments with three requested Luna 6 Max subprocesses covering the visual pipeline, source-resolution/UV audit, and runtime lighting. Generic glTF geometry simplification collapsed source atlas interpolation on broad walls and floors. Both environment builders now use the shared optimizer policy with simplification disabled and the automatic texture pass disabled; Meshopt compression, joining, material reuse, instancing, and the separate full-resolution KTX2 encoder remain enabled.
+- Repackaged all 30 visuals from their converted static scenes, refreshed asset hashes/sizes and optimization reports, regenerated the 29 generated definitions, and updated Forsaken Abbey's authored asset descriptor. Collision and navigation assets were retained. The source audit links all 1,149 shipped images to their DDS originals, with no reduced dimensions or dropped mip levels. KTX2 remains ETC1S compression; retaining source resolution does not claim lossless pixels or invented higher-resolution originals.
+- The final audit expands all GPU instance TRS and compares all 30 rebuilt environments with their original GLBs extracted from Git revision `6114a2295f14aa20a8444c61a960faeb88ac70c5`. Rank-collapsed mapped area fell 97.7% on upward floor-class surfaces and 99.5% on wall-class surfaces. These metrics include non-playable caps and decorative geometry; they measure import mapping damage rather than a percentage of visibly defective playable floors. The compact before/after reports retain texture provenance, aggregates, and selected spatial hotspots; the earlier unexpanded snapshot is archived under ignored `artifacts/rom-environment-quality-audit/` for provenance.
+- Added explicit RoM runtime lighting profiles before Bite-Sized derivation, with restrained fill for Barren Caves and HOTO. The shared directional light now targets the entrance in Standard mode and follows the camera in High mode. These profiles are runtime approximations; the static mesh importer does not restore original WDB light components.
+- The isolated Edge review loaded all 30 shipped environments without page errors and captured previous/revised entrance lighting plus revised middle/final available boss viewpoints and reverse entrance. Reviewed all six contact sheets plus the Hall of Survivors entrance at full size: the former striped floor now shows its stone atlas tiles, and the broad floor/wall mapping defects visible in the earlier captures are substantially improved. Evidence is under `artifacts/playtest/rom-environment-quality/`, including `results.json`, 150 viewpoint captures, contact sheets, `khronos-validation.json`, `package-verification.json`, and `hall-of-survivors-comparison.png`. The one initial Aeternal Circle navigation timeout passed its isolated rerun; final results contain 30 loaded entries and zero errors.
+- Targeted diagnostic captures for Grafu Castle, Treasure Trove, and The Origin are also in that directory. Grafu's flagged upward ceiling-cap group has a normal visible tiled floor above it. Some Treasure Trove/The Origin boss or hotspot anchors place the diagnostic camera outside or below the imported geometry. Surface-normal/UV metrics identify review candidates rather than proving a visible playable-floor defect, and these captures do not certify traversal, collision, combat, or a full dungeon clear.
+- The initial Sand Vortex review correctly found no terrain ROS but did not inspect the WDB's native terrain tables, so its flat-arena gap statement was incomplete. The native WDB terrain recovery below supersedes that claim; residual atlas metrics remain candidates for review rather than proof that every polygon is visibly defective.
+- The normal production graphics workflow entered full Stormwind Stockade and Forsaken Abbey, with default 150% rendering and High 200%/16× filtering/shadows. It passed with zero page errors; reviewed `artifacts/playtest/graphics-options/rom-high.png`, showing the revised Abbey wall/floor lighting. The native RoM character pack was not installed for this check, so the party uses its normal fallback bodies. Refreshed local content version `content-60e881beda3ca768`; no external deployment was performed.
+- Focused environment validation passed: 23 tests across lighting coverage, light targeting, graphics/material preparation, production loading, RoM registration, and WoW environment isolation; actual optimizer CLI and instance-transform/UV-rank regressions also passed. All 30 rebuilt GLBs passed Khronos validation after known unsupported BasisU/instancing tooling messages were filtered. Existing KTX2 and environment draw-call budgets passed without changing their baselines. A forced TypeScript build cleared a stale cached diagnostic after correcting an unrelated in-progress test import; the production Vite build, static compression, and release-bundle verification passed. The updated enemy-source-aura integration fixture subsequently passed all three tests. The release candidate check passed, including untracked files.
+
+## 2026-10-05 — Sand Vortex native terrain recovery
+
+- The authoritative `dgn_sand_vortex.wdb` contains native `header`, `chunks`, `chunk_iterations`, and `texture_list` tables. Its complete terrain is 32×32 chunks of 64×64 cells at 7.5 source units per grid interval (480 units per chunk), spanning X/Z −7,680 to +7,680. Recipe encounter bounds select chunks X/Z 14–16, nine chunks covering [−960,480] on both axes.
+- `scripts/runewaker-pipeline/sand-vortex/extract-native-terrain.mjs` now reads the source tables and writes reproducible geometry/material payloads under ignored `runewaker-export-work/sand-vortex/native-terrain-source/`. Every selected chunk has a 65×65 heightfield, 8,192 source triangles, 256² DXT1 weight maps, and source-derived visibility flags. Index order matches the native even/odd two-cell terrain topology; masked source faces retain their slots as degenerate triangles. This selected area contains no masked triangles or water layers.
+- Original Viper Canyon tile DDS files are copied without editing. Chunk layer IDs map to dirt02/24, dirt04/15, rock03/12, dirt03/20, and dirt01/10 (tile scale per terrain `texture_list`); each layer DDS is 256² DXT5. Weight-map RGB channels map to consecutive layer-ID triples. The extractor preserves the source DXT1 bytes, writes DDS wrappers for inspection, and decodes linear RGB PNGs for the runtime texture carrier; it does not invent or bake a large terrain atlas.
+- Native normals come from the serialized vertex-data RGB decoded by the original terrain shader as `(channel/255 - 0.5) * 2`; the high byte remains the visibility flag. Coordinates stay in source left-handed Y-up space in the buffers. Weight UVs are normalized local XZ with no offset or row reversal; the native shader applies tile UV scale `[tileScale,-tileScale]`, samples both splat maps at the raw normalized UV, and normalizes their RGB weights.
+- The GLB runtime adapter carries source tile and weight maps through material variants, blends them per pixel, and relies on the shared RoM dungeon lighting palette. Those palettes remain runtime approximations because the static import does not contain the original light components.
+- The focused extractor/parser suite passes all four tests, covering datastore offsets, native topology/masking, source normals and layer IDs, chunk selection, and the DXT1 wrapper. The extraction regenerated all nine chunks; all manifest paths resolve, all serialized chunk origins match the WDB header grid, and no payload bytes remain unparsed. `npm run release:check` passed at the integration checkpoint. The isolated browser review of the rebuilt package loaded Sand Vortex with zero page/console errors (7 draw calls; 56.4 FPS in the one-second preview sample). Entrance and Sandlord views show continuous native sand/cliff tiling with no obvious stretching or chunk seams; revised warm lighting makes the ground and rock faces more legible, with prop shadows still visible. The reverse entrance preview faces a nearby slope that fills the frame, so that camera angle alone does not establish route clearance. This asset preview does not certify movement, collision, or combat.
+- A post-build production browser entry on local content `4d89d2885523b400` also loaded the registered Sand Vortex with zero page errors. It used 200% render resolution, 16× filtering, High shadows, and High ambient fill (default comparison: 150%). The normal entrance view shows the native material blend beneath the party, a continuous floor-to-slope join, and visible terrain-object shadows. The party is at 0 m explored; its lower legs are partly covered by the action bar, so this image is not conclusive proof of exact foot contact. The authored entrance projects to the retained navmesh within 2.493696 m (2.45 m horizontal, +0.48 m vertical). This was an entry-only visual check: no route traversal, collision certification, or full clear was performed. Captures are in `artifacts/playtest/rom-native-terrain-entry/`.
+
+## 2026-10-06 — RoM native ground, support floors and navigation follow-up
+
+- Local content now includes Origin's full 256 native terrain chunks and 126
+  decoded water layers. Native floor sampling, heightfields, topology masks,
+  normals, splat weights, DDS dimensions and all 693 ROS placements are retained.
+  Source-scaled side projection removes the verified steep-cliff streaks while
+  preserving native XZ mapping on flat floors. Paired cliff/water screenshots
+  ran at approximately 60 FPS with no browser errors in
+  `artifacts/playtest/rom-continuation/origin-full-ground/`.
+- Water keeps all 30 native 256² normal frames and source octave settings.
+  Material clones share animation phase and retain separate transform uniforms.
+  Compatibility UV0 carriers satisfy the glTF variant contract; the shader
+  continues to sample world coordinates. Original SH lighting, projected
+  reflection/refraction and native effect type 4 remain unsupported, so this
+  water rendering is an explicit approximation.
+- Gameplay collision now honors original entity collision options and subtree
+  gates before Trekker's water/foliage exclusion. Origin uses 865,100 eligible
+  static ROS triangles and 2,094,713 visible terrain triangles. Its 24,612-polygon
+  Detour graph retains actual directed links: entry854 reaches Life Leecher;
+  Falynum has native floor but remains disconnected. No bridging links were
+  invented. Two packaged ordinary walking legs passed (2.79 m, 3/3 grounded,
+  full health, zero browser errors), recorded in
+  `artifacts/playtest/rom-continuation/origin-packaged-gameplay/results.json`.
+- Corrected support geometry uses the same inverse-placement/reflected-Z bridge
+  as native geometry. Aeternal omits only the source-proven native entry pad;
+  151 other supports remain. Its explicit all-components policy retains 1,327
+  navigation triangles and disconnected islands. Demon retains 186 supports;
+  Mystic retains 250 and uses its actual native floor DDS at measured 22.18
+  pixels/m and entry UV phase. Both textured support styles preserve native UV0
+  and use full-rank UV1 on top and side faces. These remain derived fallbacks.
+- Packaged ordinary UI/WASD entry checks completed three legs each in Aeternal,
+  Demon and Mystic (9.74 m, 7.36 m and 9.65 m); all 25 samples stayed grounded
+  with no browser errors. Standard/High screenshots are in
+  `artifacts/playtest/rom-continuation/support-packaged-gameplay/`. These checks
+  used the normal WoW healer/Ready NPC entry flow; automatic nearby combat and
+  experience gain occurred. They do not certify full traversal or combat balance.
+- The final 30-dungeon audit matches 1,615/1,615 source textures with no
+  downscales or cap hits. This summary run checks source identity/dimensions;
+  pixel fidelity is recorded separately by the codec and native-water proofs.
+  All 162 shipping dungeon GLBs pass validation with zero actionable warnings
+  or errors under the raised issue limit. Native alpha joins pass 54/54 exact
+  rows across 13 source hashes. The 39 generic pixel-role suggestions match
+  the old baseline and are not native alpha evidence; tree04/tree08 KTX cutoffs
+  retain their verified native policy.
+- Regeneration is owned by `build-environment.mjs`, native terrain/water
+  builders, `assemble-native-collision.mjs`, and
+  `build-native-core-support-candidate.mjs`. Strict query cache identity includes
+  source and extractor hashes. The three primary/support producer smoke tests,
+  bounded collision chunk test, focused scene/navigation tests and TypeScript
+  project build pass. Existing unrelated repository changes are preserved.
+- Final source/UV-compatible Origin and textured Mystic were checked again
+  through `build-final-20261006`: two Origin legs (2.58 m, 3/3 grounded) and
+  three Mystic legs (9.88 m, 9/9 grounded), with no browser errors. The final
+  Mystic High screenshot confirms that the former gray pad matches the native
+  entry floor pattern. Evidence:
+  `artifacts/playtest/rom-continuation/final-packaged-gameplay/`. Final production
+  build, precompression, bundle verification, KTX2 audit and performance audit
+  pass. The local content version is `content-2160be7ff070205b`.
+
+## 2026-10-05 — RoM texture codec and native material follow-up
+
+- Upgraded sixteen reviewed texture images across nine packages to UASTC level 2 with Zstandard 18 and no RDO. The decoded source comparison improved RGB PSNR from 20.34–26.05 dB to 48.43–57.31 dB, with image dimensions unchanged. The exact basename policy avoids accidentally upgrading similarly named roughness maps; other images retain ETC1S compression. Source/pixel evidence and the frozen codec baseline are in `artifacts/rom-environment-quality-audit/codec-upgrade/`.
+- Restored native-supported normal transparency on The Origin's `act_precipice_01`, `act_precipice_02`, and `act_waterfall02` resources, plus native cutout thresholds where the source image actually contains alpha. All 693 WDB placements, 220 resources, original bounds, geometry, UVs and primary material modes remain intact. The four affected placement matrices differ from the source plan by at most 6.4×10⁻⁶ due to Float32 storage. The native alpha provenance is shipped in `src/assets/game/dungeons/the-origin/native-material-alpha-report.json`. Five source additive wave material rows remain unsupported and unchanged; their mode was not inferred from filenames.
+- Fresh Origin preview loading passed with zero page/console errors after one startup navigation timeout passed its rerun. Additional cameras around waterfall hotspot cell `[1,-1,8]` also loaded cleanly, but nearby decorative geometry and opaque foliage cards obscure these views, so the screenshots do not certify the waterfall's full visual appearance. The source report, identical embedded RGBA, decoded KTX2 alpha range 0–173 (MAE 1.365), and material mode evidence support this scoped alpha correction. Full native effects, other foliage cutouts, and the existing entrance projection review remain separate source/rendering follow-ups.
+- The final source audit covers 30/30 environments and 1,182/1,182 embedded textures with matching source DDS dimensions: zero downsized images, zero 2048-cap hits, zero dropped mip chains, and all 1,182 pixel comparisons completed. Sand Vortex now has 33 embedded textures and no untextured terrain fidelity gap. Native terrain UV diagnostics measure the default layer carrier; the runtime blends five source-scaled layers, so a single-layer density metric does not describe the entire splat shader.
+- All 30 package asset checksums/sizes match their metadata; all 155 package GLBs pass Khronos validation after known unsupported BasisU/instancing tooling messages are filtered. The updated review results contain 30 loaded environments and all six contact sheets were refreshed. The final production build, static compression and release-bundle verification passed; KTX2 has no fallback textures and environment performance has no regressions. Local content is `content-e261b88ce9be26fa`; no external deployment was performed. The audit retains source-authored UV residuals as review candidates and does not certify every polygon, route or encounter.
+
+## 2026-10-05 — CoA Decapitate strike integration and Guardian source audit
+
+- Corrected Decapitate 804414/node 5744 from two imported damage events to one
+  normalized weapon strike using the captured Vol'Jin tooltip: flat 27 plus
+  220% weapon damage including normalized AP. Main-hand item classification
+  supplies 3.3/1.7/2.4-second normalization; actual speed does not classify the
+  weapon. Players, companions, and their scorers share the projection. Unknown
+  classifications retain the imported fallback. The generic AzerothCore
+  flat-bonus ordering differs; the custom CoA script is unavailable, so the
+  explicit tooltip governs this narrow interpretation. Extra-Energy conversion
+  remains unfinished; existing health gates, payment, and Beheader refund pass.
+- Focused deterministic fixtures prove one correctly calculated damage event
+  through actual player and normal NPC casts. The focused set passes 22 tests
+  across five files; Ready-five compatibility passes seven Vitest and twenty
+  Python tests. No catalog records or readiness pools changed.
+- The selected five builds were tested at level 60, seed 1, for at most 30
+  seconds in isolated headless boss rooms sourced from Cellblock Rush's
+  catalog-template Hamhock, `bite-sized-stormwind-stockade/entry-1717`. Guardian
+  and Witch Doctor seats time out with 80,007/77,898 boss health; Barbarian
+  defeats him in 12.8 seconds; Ranger times out with 29,396; Pyromancer defeats
+  him in 24.8 seconds. Every seat has zero deaths. The final Barbarian scorer
+  chooses other attacks and records no Decapitate cast; separate actual-cast
+  regressions establish the strike behavior. Low pressure prevents tank/healer
+  completeness or balance conclusions. This is not a live dungeon clear or
+  browser visual check. A fresh live Cellblock Rush run can select another boss.
+  Raw configs/results are preserved under ignored
+  `artifacts/coa-first-party/stockade-decapitate-2026-10-05/`; policy is
+  `playerbots-inspired-v11-coa-decapitate-strike`. All five remain Needs Testing.
+- The Guardian audit establishes trainer metadata for rankless Tower/Line and
+  why they need an explicit import path and formation-specific modifiers.
+  Paragon's secondary stack text and child damage/buff rows support the August
+  Heavy Blow/Paragon Smash intent; its stale primary Tower/Bastion description
+  remains recorded. Exact cached IDs/fields and acquisition boundaries are in
+  `docs/COA_GUARDIAN_FORMATION_AUDIT_2026-10-05.md`. This audit adds no actions or
+  assumed formation state. The canonical CoA vault note is updated separately.
+- Final production build, TypeScript, asset audits, static compression and
+  release-bundle verification pass. `npm run release:check` passes, including
+  untracked files. No external deployment was performed.
+
+## 2026-10-05 — Ordinary NPC spell blocking and Cellblock Rush default
+
+- Routine live QA now defaults to **Stormwind Stockade: Cellblock Rush**
+  (`bite-sized-stormwind-stockade`). Shared harness defaults, contributor rules
+  and game/vault workflows record its one seeded boss. Historical full Stockade
+  evidence remains intact; new full-source runs require explicit selection.
+- Source-bound direct physical player and companion spells now use ordinary
+  NPC defenses and a separate post-hit block roll. Exact selected-rank metadata
+  determines eligibility. Critical hits can also block, and outgoing events
+  retain block flags/amounts. Native RoM, magical, periodic and unbound packets
+  stay outside this phase. Non-direct full-hit blocks and source aura imports
+  for ignore-hit-direction/critical-block doubling remain separate work.
+- Ten directly affected Vitest files pass 134 checks. Actual Heroic Strike and
+  generated Hunter Steady Shot fixtures prove crit-plus-block with labeled
+  deterministic setup. Ready-five passes seven Vitest/twenty Python checks;
+  the final 8 GB production build/hooks and subsequent TypeScript check pass.
+- Normal offline level-30 Discipline Priest entry with four Ready companions
+  clears actual seeded Hamhock (`creature-79098`, `entry-1717`) in 03:18.0 on
+  Easy/Normal. Native required/killed IDs agree 1/1. No player/party death or
+  resurrection recovery is observed; all five actors finish alive. Nine
+  effective heals total 2,801.9 HP. The results roster naturally reaches player
+  level 32 and companion level 31; cap 80 remains unchanged.
+  The UI-created profile was advanced to level 30 through isolated shell
+  progression setup before entry; no combat-state writes were used during play.
+- Run `playtest-artifacts/stockade-spell-blocking/20261005-202545227` captures
+  830 unique sampled events and 29 block indicators. Six party-to-hostile
+  blocks include two Revenge and two Sinister Strike hits with event-level
+  amounts; a correlated Revenge result explicitly records its post-hit block
+  roll. Four hostile `spell-15657` blocks use the existing companion defender.
+  No live crit-plus-block is observed. Read-only latest-200 ring snapshots can
+  miss events; this is scoped smoke evidence, not balance certification.
+- Root reviewed desktop/compact roster, entrance, route 50/75 and the native
+  one-boss result screenshots. Bodies, staff, shoulders and visible companion
+  weapon/shield silhouettes render. Close-party overlap and edge occlusion
+  persist; compact roster text is clipped, though normal Continue hit testing
+  passes at both sizes without a forced click. These captures do not certify
+  every animation or companion silhouette.
+- Frozen distribution aggregate SHA-256 is
+  `864781d16c3c0a26457b6acd7a9bd31357116f856e6b27258b5b633d2f224fac`
+  (5,964 files, 116,755,437 bytes), unchanged before/after. Full content manifest
+  JSON is retained at both boundaries and stays unchanged; all 2,124 served
+  objects match their hash addresses and manifest references. Page errors are
+  zero; the sole console item is a favicon 404. The first entrance-only failed
+  harness attempt remains preserved: distant boss loading required reading
+  actual run selection rather than assuming its mob existed at entry.
+- Source/runtime report: `docs/NPC_SPELL_BLOCKING_2026-10-05.md`. Exact live
+  provenance, sampling boundaries and cleanup:
+  `docs/STOCKADE_SPELL_BLOCKING_PLAYTEST_2026-10-05.md`.
+
+## 2026-10-05 — RoM walking-route and native-material pass
+
+- Sand Vortex now blends source-scaled side projections on steep terrain using
+  the original smooth normals, layer scales, and splat weights. Native XZ
+  sampling remains on level ground. Same-camera browser comparisons compile
+  without errors and remove the vertical cliff streaks; the sampled flat-floor
+  region is pixel-identical across all 630,000 RGB channels. No terrain geometry,
+  collision, weight images, or source DDS dimensions changed for this fix.
+- Corrected entrance heading conversion through inverse primary rotation and
+  component scale, source handedness, and normalized XZ. Six rotated WDB
+  assemblies receive updated local forward/yaw; authored Lua directions remain
+  in provenance. The focused assembly suite passes all seven checks, including
+  nonuniform and negative scale.
+- Normal offline Sand Vortex entry and ordinary WASD reach all four waypoints
+  from the entrance to Sandlord, including the gentle slope transition. The
+  route records 62.96 m explored, zero ungrounded samples, full player health,
+  and zero browser errors. Standard/High captures show native floor and cliff
+  detail along the route. The final boss camera is partly occluded by a tree;
+  telemetry establishes movement but that capture does not certify the boss
+  view. No teleport, combat-store mutation, or full clear was used. Evidence:
+  `artifacts/playtest/rom-visual-routes/sand-route-final/` and
+  `artifacts/rom-environment-quality-audit/sand-vortex-cliffs/`.
+- Ordinary UI entry and Standard/High options loaded the other 28 RoM dungeons
+  with zero browser errors. All entries settled grounded and all final settings
+  recorded High shadows, 150% resolution, 16× filtering, and Authored fill.
+  Reviewed all six paired contact sheets under
+  `artifacts/playtest/rom-visual-routes/all-28-lighting-v5/`. Floor/wall detail
+  remains legible across the dungeon palettes; some entrance cameras face a
+  wall or crowded party. Existing untextured source-anchor support surfaces in
+  Aeternal Circle remain a derived-geometry fidelity gap. Demon Stronghold's
+  base package received the reviewed native-floor support fallback below.
+  These entry captures
+  do not establish complete route clearance or combat balance.
+
+## 2026-10-05 — The Origin native ground and runtime-entry review
+
+- Recovered authoritative WDB terrain chunk `[2,2]` from `dgn_the_origin.wdb`
+  for the entrance-path region. The source 64×64-cell topology keeps all 8,192
+  triangle slots, source normals, visibility, four tile DDS layers/scales, and
+  two DXT1 splat maps. Of its 8,192 source faces, 8,146 visible faces are
+  packaged for collision/nav; 46 masked faces remain excluded. One 30,332-byte
+  water-layer payload is left unparsed, so the package does not claim native
+  water rendering or collision.
+- The old Lua spawn `[828,263,963]` and direction `343` stay in provenance, but
+  its projected gameplay position was trapped against WDB geometry. Runtime
+  entry now derives from source path graph 1026 node 854 and faces toward source
+  node 80 as a heading reference. The source node center is 1.936 m from the
+  reviewed nav point; runtime sourcePosition is the exact inverse transform of
+  that point, not the unprojected node center. Node 798 was grounded but slid
+  against blocking geometry with an obstructed camera and was rejected.
+  Recast included the 15 existing WDB collision files plus the native terrain
+  collision and produced 98,298 triangles. The reviewed node854 entry passed
+  two short ordinary walking legs: 2.846 m explored, 3/3 grounded samples,
+  both targets reached, full health, and zero browser errors. Standard/High
+  captures show clearer native ground and columns. `gameplayReviewed` and
+  `reviewRequired=false` certify only this short entry check; the HUD still
+  reports party route blocked farther inside.
+- Registered the terrain collision separately and refreshed the Origin nav GLB
+  and checksums. The parser suite passes 5/5, heading/assembly suite 7/7, and
+  all packaged collision/nav file sizes and hashes match import metadata.
+  Evidence: `artifacts/playtest/rom-visual-routes/origin-854-final/`.
+
+## 2026-10-05 — RoM source texture and material follow-up
+
+- Barren Caves' broad `wretched_tunnel` rock material now repeats four times
+  through a separately retained UV channel. Exact ROS/material/texture hashes
+  gate the recipe policy; all 31 uses receive the same adaptation to preserve
+  continuity. The canonical source retains UV0, while shipping optimization may
+  prune or renumber unused UV sets. Route, floor, and wall captures of the actual
+  package load without errors and show finer rock detail. The 30-dungeon hotspot
+  ledger records the remaining source caps, trim, effects, and reviewed atlases;
+  Pasper's atlas is intentionally unchanged. Evidence:
+  `artifacts/playtest/rom-environment-quality/rom-texture-pipeline-triage/`.
+- Origin source-proven material corrections restore additive water waves,
+  prelit waterfalls, foliage cutouts, and cocoon transparency. Thirteen native ROS
+  resource reports are joined by exact source hash and material identity; five
+  additive materials and six prelit materials are recorded separately. Prelit
+  materials lose unused specular metadata before optimization, avoiding an
+  incompatible glTF extension pair. The two additional cavern plant materials
+  use native MASK threshold 192/255; the opaque cavern surface remains opaque.
+- Five additional tree resources across nine placements restore native MASK
+  thresholds 120/255 or 192/255. The prior tree04/tree08 shipping textures had
+  lost alpha during opaque compression. Re-encoding after the material fix
+  restores actual 1024×512 KTX2 alpha from 0 to 255: tree04 has 273,648 pixels
+  below its cutoff, tree08 has 302,448. No source image or dimension changed.
+- Final Origin source verification preserves all 2,696 geometry accessors,
+  481 embedded images, 693 ROS placement matrices, and the appended native
+  terrain. The other nine cavern primitives retain their material state.
+  The isolated shipped giant-tree check confirms cutout pixels on the actual
+  KTX2. Evidence: `artifacts/rom-environment-quality-audit/origin-followup/`,
+  including `canonical-final-integrity-proof.json` and `final-provenance/`.
+- Final thirteen-resource evidence is in `thirteen-final-provenance/` and
+  `remaining-foliage/final-integrity-proof.json`; all 2,696 original accessors,
+  481 images, and 693 ROS matrices remain identical. The shipping visual is
+  29,960,084 bytes, SHA-256
+  `528ac316d39f1ca13f9bb93ac170a585fd9b0ef87ff140183452fcd99bb5d145`.
+- Base Demon Stronghold's 186 generated support boxes now sample native
+  `entrance_floor_02` at the measured 36.24 px/m entrance scale and phase.
+  All 2,232 box triangles have noncollapsed UVs, including vertical faces;
+  all 826 original accessors and 56 images are preserved. Exact source hashes
+  guard this opt-in, idempotent pipeline policy. It remains a derived support
+  fallback, not reconstructed native floors at every anchor. The level125
+  variant is unchanged. Evidence: `demon-support-integrity-proof.json` under
+  `artifacts/playtest/rom-environment-quality/rom-texture-pipeline-triage/`.
+- The refreshed full audit compares all 1,188 shipped textures against source
+  dimensions and decoded source PNGs: zero downscaled textures, zero 2048-size
+  cap hits, and zero audit issues. All 30 package hashes/sizes match, and all
+  156 packaged GLBs pass Khronos validation with zero actionable errors or
+  warnings after the documented BasisU/instancing tooling exclusions. Focused
+  pipeline tests pass 27/27, focused rendering tests pass 64/64, and native-alpha
+  Python tests pass 3/3. These checks
+  preserve source resolution and material intent; they do not certify a full
+  dungeon clear or native WDB light placement.
+- Final local content publication is `content-51e1d8717914b459`. Prebuild,
+  TypeScript, Vite production build, Brotli/gzip precompression, and release
+  bundle verification pass. The shared `dist/assets` directory is locked, so
+  the verified build is `artifacts/playtest/rom-visual-routes/build-final-20261005`.
+  Final normal-UI evidence in `final-packaged/` records Demon Stronghold's
+  9.795 m leg reached with 9/9 grounded samples and Origin's two entrance legs
+  reached with 2.531 m explored and 3/3 grounded samples. Both have full player
+  health and zero browser errors; Standard/High screenshots were reviewed.
+  Origin foliage cutouts are visible, while nearby native columns/architecture
+  still crowd its entrance camera and the party route remains blocked.
+  Aeternal's untextured derived supports and Origin's unparsed water payload
+  remain unresolved. These runs do not certify full dungeon traversal.
+
+## 2026-10-05 — CoA Guardian formation/Paragon integration
+
+- Integrated Tower/Line learning and owned cast projection, current shield armor
+  and Footman's Calling mitigation, Line movement/root duration, Paragon stacks,
+  ready Heavy Blow/Smash projection, block charges and Brace cooldown reduction.
+  Healthy Vanguard companion scoring prefers a ready Smash over routine upkeep;
+  existing low-health mitigation and urgent reactive choices retain priority.
+- Live entry exposed and fixed free startup seeding of both raw formations and
+  removal of a saved learned Heavy Blow binding. Three precombat regressions,
+  a legal talent-allocation entry regression and the separate binding allowlist
+  check cover those failures. Unique focused coverage totals 66 tests in 15
+  files; Ready-five compatibility passes 7 Vitest + 20 Python, avatar checks 56,
+  and narrow shared enemy/boss-guide compatibility checks 17.
+- Prebuild/asset checks pass. The shared dist/assets directory was locked at
+  rebuild, so final TypeScript/Vite, compression and bundle verification use
+  artifacts/coa-first-party/stockade-guardian-2026-10-05/dist-final. The reusable
+  harness records its confined --dist path and source-derived public preload
+  aliases; no production deployment or readiness-pool change is implied.
+- Completed an isolated 180-second Cellblock Rush browser diagnostic with
+  Guardian/Vanguard, Witch Doctor/Brewing, Barbarian/Brutality, Ranger/Archery
+  and Pyromancer/Incineration. The exact roster and legal Paragon variant are
+  configured before normal Tank entry, then normal action-bar/movement input
+  casts Tower -> Line -> Tower. Entry has no seeded form; Heavy Blow survives
+  in R2. Base formation cost 20 resolves to 17 for this selected talent build.
+- Fresh run selected Dextren Ward (entry-1663, creature-79116), confirmed by the
+  intro card. The route remains near the entrance: about 8 m explored, 5.74 m
+  net movement, full health, no observed boss actor, no successful Heavy Blow
+  cast or natural Paragon proc, and no healing pressure/clear claim. Regression
+  fixtures separately exercise actual Smash/block/Brace behavior. All five
+  remain Needs Testing; a selectable first-party UI preset and a navigated
+  combat playtest are still needed.
+- Visually reviewed the roster and in-dungeon captures. Guardian's body, shield
+  and weapon silhouettes are present; clustered companions and temporary gold
+  effects limit a full party gear inspection. Screenshot 02 records the intro
+  countdown rather than the body; 03/04/06 show the dungeon. Heavy Blow is visible
+  at the endpoint. Page exceptions: 0; console 404: 1 unclassified; captured
+  failed requests/HTTP errors: 0. Evidence is ignored under
+  artifacts/playtest/coa-guardian-formations/20261006-011259958/; exact source
+  limits and results are in docs/COA_FIRST_PARTY_MILESTONE.md and
+  docs/COA_GUARDIAN_FORMATION_AUDIT_2026-10-05.md.
+
+`npm run release:check` passes, including untracked files. The Git index is
+unchanged; no commit, deployment or readiness promotion was performed.
+
+## 2026-10-05 — Enemy weapon utilities and Cellblock Rush
+
+- Integrated source-bound Bloodlust, Disarm/Infected Wound and explicit
+  Dual Wield/Thrash through separate GPT-6 Luna extra-high subprocesses.
+  Focused coverage passes 360 distinct Vitest, five Node and twenty Python
+  cases, including Ready-five and 56 avatar checks. Cap 80, readiness and loot
+  scope remain unchanged. Source limits and remaining parity are recorded in
+  `docs/ENEMY_WEAPON_UTILITIES_2026-10-05.md` and the mechanics backlog.
+- Standard prebuild checks and 8GB TypeScript pass. The default Vite output
+  hits Windows EPERM clearing shared dist/assets. A fresh isolated production
+  output passes Vite, precompression and release-bundle verification, then its
+  complete distribution and manifest-referenced content are frozen for QA.
+- Run `20261006-012123000` uses normal Discipline Priest level-30 healer entry
+  and four Ready NPCs. The fresh run selects Hamhock (`entry-1717`,
+  `creature-79098`) and clears Cellblock Rush's one native objective in 3:14.6.
+  There are no observed player or companion deaths. Normal combat advances
+  the player to 32 and companions to 31 by the result screen.
+- Root visually inspected roster, equipped dungeon silhouettes and results:
+  the player body, robe, shoulders and staff are visible. Route-75 also shows
+  companion bodies and staff/club/shield silhouettes; clustering and overlap
+  limit a complete per-companion gear inspection. The results
+  capture explicitly shows Hamhock and Bosses defeated 1/1. There are no page
+  exceptions and one favicon.ico 404.
+- Build aggregate `58dfe50765b0635caeae50bd3eca2fd079cd69cce0e571ff6cb2dd5aeec77b39`
+  and content manifest `36614a23f3b98b4289bc3726c8a9bdb075445ada50dfce6504ade4e36325be1f`
+  are unchanged before/after. Root independently rehashed all 2,124 served
+  manifest-referenced objects and checked the actual loaded entry, combat and
+  runtime chunks. Evidence: `playtest-artifacts/stockade-weapon-utilities/20261006-012123000/`,
+  including `root-evidence-audit.json`.
+- The first harness checkpoint reducer omits its new weapon fields, making
+  all-null summarized state a harness gap. The corrected recorder retains
+  those fields and globally samples the three source utility aura IDs,
+  including player and party recipients. Syntax preparation passes.
+- Corrected run `20261006-013934624` uses the same frozen inputs, selects
+  Bazil Thredd (`entry-1716`, `creature-79124`) and verifies the native 1/1 clear
+  in 3:20.7 with no player/party deaths and 1,504.5 effective player healing.
+  All 328 mob snapshots expose weapon fields; 35 show Bazil's Dual Wield
+  capability enabled. All sampled offhand timers are zero; hand-specific
+  damage and nonempty Thrash queues are unobserved. These are capability
+  snapshots, not proof of 35 activations or any offhand hit.
+- Global store evidence retains 26 snapshots of the tank's Wound instance,
+  covering four distinct source/application-time tuples. Each is a one-stack
+  five-minute Disease aura. Bloodlust and Disarm auras are unobserved in this
+  retest; deterministic imported-runtime tests cover them and queued Thrash.
+- Root independently checks the unchanged hashes and all 2,124 served objects
+  again and visually reviews the new roster, dungeon and results captures.
+  The player body and equipped silhouette remain present; clustered companions
+  limit a full gear audit. Results explicitly show Bazil Thredd and 1/1.
+  There are no page exceptions and one favicon 404. Evidence and root audit:
+  `playtest-artifacts/stockade-weapon-utilities/20261006-013934624/`.
+- Final `npm run release:check` passes, including untracked files. Game and
+  vault diffs are reviewed separately; no stage, commit or deployment is made.
+
+## 2026-10-05 — Weapon parity follow-up and combat-time recording
+
+- Integrated player/party source Disarm aura and selected-hand contribution
+  views, verified native active-five unarmed inputs, intrinsic enemy offhand
+  flags, and actual swing hand/source metadata. A review regression also removes
+  default-speed AP contribution from disabled offhand/ranged channels while
+  preserving independent scripted AP. Mismatched backup equipment tables remain
+  unbound; native custom gear, rate and prerequisite limits stay explicit.
+- Final focused checks pass 102/102; Ready-five passes 7 Vitest + 20 Python;
+  avatar checks pass 56/56; recorder checks pass 3/3. The standard production
+  build and release-bundle pipeline pass. No full suite/all-class sweep ran.
+- The corrected frozen build SHA is
+  `fcdde36bbd63ac711453a57562a38d7a1aa7ff2e8cafd4acb0e65a7a4dc27efe`.
+  Its unchanged isolated content manifest SHA is
+  `ac0fa2a867ba0a495c172de2b89532d69b32d56d032482b26020ce2d189f074b`.
+  All 27,571 referenced objects were verified in the original freeze; the final
+  freeze reuses that copy, rechecks the manifests and hashes the corrected build.
+- Final Cellblock Rush run `20261006-035524177` seeds Hamhock (1717,
+  `creature-79098`) and clears exactly 1/1 in 03:15.1. The UI-created Discipline
+  profile receives isolated pre-entry level-30 preparation, then uses normal
+  controls with four Ready NPCs and read-only telemetry. No Boss Testing or
+  combat-state writes. No recorded deaths or runtime exceptions.
+- The committed-state recorder retains 842 consecutive events, zero gaps,
+  errors or overflow; 73 source main-hand results; 528 weapon-state observations;
+  and 15 Bloodlust, 3 Disarm and 49 Wound snapshots. Party debuffs are included.
+  This seed has no observed offhand/Thrash result; scoped runtime fixtures cover
+  them. Historical first-build run `20261006-034834217` separately clears Bazil
+  and records 44 main-hand results plus one blocked offhand result, positive
+  hand timers and party Disarm, before the final AP-speed correction.
+- Root views final roster, dungeon equipment, mid-route and native results
+  captures. Player/party bodies and staff/shield/weapon silhouettes remain
+  visible; clustering limits a complete gear audit. Hamhock and 1/1 are readable
+  in results. The only outer browser console item is favicon 404.
+- Full report: `docs/WEAPON_PARITY_FOLLOWUP_2026-10-05.md`. Final evidence:
+  `playtest-artifacts/stockade-weapon-utilities/20261006-035524177/`. Historical
+  first-build evidence stays separate. Level cap 80, loot and readiness unchanged.
+
+## 2026-10-06 — Enemy source parity and pinned equipment
+
+- Implemented source NPC block inputs/critical-block results, separate victim
+  and area shout targets, actual player/party hand evidence, and recovered
+  pinned spawn-equipment mappings. Validated offhand weapons enable ordinary
+  NPC AUTO attacks with independent timers and reset persistence. Synthetic,
+  mismatched, empty, shield and native assignments fail closed. No loot wiring.
+- Focused source/hand tests pass; the final affected equipment/runtime/population
+  group passes 73 tests. Node importer/compiler/recorder checks pass 20;
+  Ready-five passes 7 Vitest + 20 Python; avatar checks pass 56. The standard
+  TypeScript/Vite, asset-audit, compression and release-bundle build passes.
+- One final frozen Cellblock Rush run `20261006-054628643` selects Targorr
+  the Dread, entry 1696 / `creature-84027`, and clears native 1/1 in 03:10.3.
+  UI-created Discipline profile with isolated pre-entry level-30 preparation,
+  four Ready NPCs, ordinary inputs and read-only telemetry. No recorded deaths;
+  2,845.8 effective player healing. No Boss Testing or combat-state writes.
+- Recorder: 794 consecutive events, zero gaps/errors/overflow, 60 NPC mainhand,
+  eight offhand and four source-3391 Thrash extra-mainhand results; 162 actual
+  player/party hand events, three committed Disarm windows and four hand results
+  during those windows. Timing/contribution fixtures remain the formula proof.
+  Disarm/Wound appear; Bloodlust/block/shout applications are unobserved here.
+- Root views roster, dungeon gear, route and results images. Character bodies
+  and staff/shield/sword/dagger silhouettes remain visible; overlap limits a
+  full companion gear audit. Boss-arrival.png captures the native results
+  overlay with Targorr and 1/1. No page exceptions; one favicon 404.
+- Build SHA `50f723e4737f01fe82973c1a9c8259a4358179c1df74373b587fa76fb69b44a7`
+  and content manifest SHA
+  `ac0fa2a867ba0a495c172de2b89532d69b32d56d032482b26020ce2d189f074b`
+  remain unchanged. The harness verifies 2,124 actually served content objects.
+  Full report: `docs/ENEMY_SOURCE_PARITY_CONTINUATION_2026-10-06.md`.
+- Independent Luna audit rehashes all 5,967 build files, seven loaded chunks and
+  2,124 served objects, and checks recorder IDs 9–802. Raw response bodies are
+  not persisted; observed URLs/status metadata match immutable local sources.
+  Audit: `playtest-artifacts/stockade-weapon-utilities/20261006-054628643/independent-evidence-audit.json`.
+  Release checks pass 399 changed paths including untracked files. Game/vault
+  diffs are reviewed separately; task vault text hygiene passes and the worklog
+  closes. All subprocesses use Luna extra-high; no staging, commit or deployment.
+
+## 2026-10-06 — WoW texture and lighting access
+
+- Three requested GPT-6 Luna extra-high workers reviewed shared graphics access,
+  lighting inheritance and imported texture quality. WoW uses the same filtering,
+  ambient-fill, shadow and KTX2 paths as RoM. A wrong-source fallback presentation
+  in randomized/custom bite-sized layouts now inherits the selected source's
+  lighting and point lights; matching tuned baselines retain their presentation.
+- The isolated asset review loads all 54 actual WoW visuals plus Cellblock Rush:
+  55/55 successful entries, zero page errors, 271 captures and ten review sheets.
+  High filtering and camera-following shadow targets pass. Root inspects the
+  sheets and detailed Stockade/Shadowfang images. These are sampled asset views,
+  not combat, traversal, collision or full room coverage.
+- The audit finds 8,099 KTX2 images with zero fallback textures. All 7,980
+  source-comparable dimensions are retained across 53 staged inputs; authored
+  Wailing Caverns lacks its matching staging source. Broad flat terrain in some
+  converted inputs remains a fidelity gap; it is not optimizer texture downsizing.
+- Safe WoW optimization now disables automatic simplification and image resizing,
+  with final Meshopt packing after KTX2. Shadowfang's source-pinned exception
+  reduces only untextured terrain/liquid. Its local installed visual is 8,566,408
+  bytes, SHA `2c562fb61596a536d728385c4f2bdfdc3230f4f84ef40875b9a568e18a82ef9d`.
+  It restores textured topology close to source while retaining every texture
+  payload and all 26 collision plus navigation hashes.
+- Final bounded Shadowfang samples retain about 60 FPS with unchanged draw calls
+  and 7.5–9.8% more rendered triangles than the prior asset. The full-source
+  candidate was rejected for nearly four times the rendered triangles. These
+  short capped Edge samples exclude gameplay actors/physics and mobile headroom.
+- Focused checks pass 118 Vitest and five pipeline fixtures. The final normal
+  production build passes TypeScript, Vite, texture/performance audits,
+  precompression and release-bundle verification. Release hygiene passes
+  including untracked paths. No commit or deployment.
+- The first Options smoke uses the last available bundle, selects 200%, 16× and
+  High shadows/fill, observes Bazil Thredd in Cellblock Rush, carries settings to
+  Forsaken Abbey and reports zero page errors. Evidence remains separately at
+  `artifacts/playtest/graphics-options/run-wow-access-2026-10-06/`; it does not
+  establish this task's new lighting fix in that old bundle.
+- The final freshly built bundle is frozen separately under
+  `artifacts/playtest/graphics-options/run-wow-current-build-2026-10-06/`.
+  Entry `index-BvFyaleo.js`, built `2026-10-06T17:39:29.274Z`, again seeds Bazil
+  Thredd, verifies all four quality controls, and carries 200% resolution into
+  Forsaken Abbey with zero page errors. Root views Options/WoW High/RoM High:
+  textured surfaces and equipped actor silhouettes remain visible. This is an
+  entry/settings smoke, not encounter validation.
+- Report: `docs/WOW_ENVIRONMENT_QUALITY_2026-10-06.md`. Asset/benchmark evidence:
+  `artifacts/playtest/wow-environment-quality/`. Full quality audit:
+  `artifacts/dungeon-pipeline/wow-visual-quality-audit.md`.
+
+## 2026-10-06 — Daze, fear movement and immunity escape
+
+- Integrated exact WoW335 Shield Bash 72 child Daze 29703, safe source-informed
+  fear movement for the player/party/mobile summons, and source-proven Berserker
+  Rage/Bladestorm purge with player/party escape. Focused checks pass 126/126
+  across 15 files; Ready-five passes 7 Vitest plus 20 Python checks. The final
+  8 GB TypeScript/Vite production build and required asset/bundle audits pass.
+- One final frozen Cellblock Rush run `20261006-172609059` selects Hamhock,
+  entry 1717 / `creature-79098`, and clears native 1/1 in 03:09.7. No recorded
+  player/party deaths; seven effective player healing windows total 2,444 HP.
+  Normal entry/inputs and four Ready companions follow isolated pre-entry
+  level-30 preparation. No Boss Testing or combat-state writes.
+- Recorder: 797 consecutive committed events, IDs 9–805, zero gaps/errors/loss;
+  70 NPC and 161 player/party hand results. Actual Shield Bash events
+  574/659/777 immediately precede accepted Daze grants 575/660/778 on Hamhock.
+  Duration/50% speed remain runtime-fixture proof, because no retained live
+  slow-state sample measures them. Source shout/fear/purge, block, offhand and
+  Thrash applications are unobserved in this seed.
+- Root views roster, equipped party, route-50 and native results images:
+  bodies and equipment silhouettes render, with close stair overlap limiting
+  individual companion gear inspection. The historical boss-arrival filename
+  contains the results overlay. Nine actual PNGs exist; declared camera-drag
+  paths have no files, so no live camera proof is credited. Zero page errors;
+  the only console error is favicon 404. The owned harness cleans up.
+- Build SHA
+  `2ded1eedcf7cfb60d53f6c82a33c088c05ff4d881bb258852c035df94770c8d1`
+  and fresh content manifest SHA
+  `85fc9f1c504c82aa794806b8a634596a07b34ba63e0bf66530243fe76e2d0356`
+  remain unchanged. All 27,587 referenced content objects were verified in the
+  fresh freeze. Root audit rehashes all 5,967 build files, seven loaded chunks
+  and 2,124 actually served manifest objects, and passes with zero failures.
+  Raw network response bodies are not retained. Historical full-clear fields
+  certify only the entered one-boss slice.
+- Full report: `docs/ENEMY_CONTROL_RUNTIME_FOLLOWUP_2026-10-06.md`. Live evidence
+  and root audit: `playtest-artifacts/stockade-weapon-utilities/20261006-172609059/`.
+  Source gaps, level cap 80, loot scope and readiness boundaries remain explicit.
+- Independent Luna extra-high evidence review confirms the seeded clear,
+  fingerprints, event continuity and three Daze grants. One timestamp pair is
+  reversed by 1 ms; consecutive event IDs establish continuity. Final game
+  release hygiene and separate task vault diff/text checks pass; memory notes
+  are updated and the worklog closes. No staging, commit or deployment.
+
+### 2026-10-06 — RoM texture, lighting and native-material pass
+
+- Luna subprocesses review all 30 base RoM dungeons through source-material
+  inspection, zone-lighting recovery and sampled runtime views. Room-camera
+  coverage includes 30 entrances, 66 boss anchors, 74 corridor samples and 36
+  mapping hotspots; diagnostic cameras do not establish walking connectivity.
+- Barren replaces unsafe whole-atlas repetition with four exact native floor
+  islands and one native wall crop. HOTO v2 uses original full stone/moss tiles
+  at two metres on true floor materials and adds source detail to cliff surfaces.
+  Original geometry, UV0 and source PNGs are preserved. HOTO v2 changes only
+  approved material metadata and retains every non-JSON GLB chunk.
+- Native WDB lighting now drives the RoM pass, with source colors, zone inheritance
+  and Standard/High shadows; Grafu retains a documented entrance fallback.
+  Source waterfall, magma, giant-tree
+  and vine alpha semantics are restored where proven; Origin water gains bounded
+  reflection/refraction capture. The report records original-engine limitations.
+- Final build: `artifacts/playtest/rom-deep-review/build-surface-v5-20261006`;
+  index SHA `ab46b9c47a0ba27b0b1c2e0a2c65ff02c613b7b87c54d37fe5ec9a7a41707e27`.
+  Frozen content is `content-d1ec6f1f858f21c1`, manifest SHA
+  `e102c30cc47a4130fff88809ce53453929e86e6f766b0afbbae8a8ad4a52fd7c`.
+- TypeScript/build/precompression/bundle checks, 83 focused runtime tests and
+  17 pipeline tests pass. All 1,620 matched textures across 30 dungeons retain
+  source resolution, including five exact crops; zero downsizing/cap hits/issues.
+  All 162 dungeon GLBs validate with zero errors/warnings. Existing visual
+  performance budgets have no regressions; 106 visual assets have no KTX fallback.
+- Final ordinary Barren/HOTO/Origin movement reaches 48/48 legs, with 89/89
+  grounded samples, 101.826 m travelled and zero errors. HOTO is only a six-metre
+  probe, and Origin follows the connected Life route. Root views the final Barren
+  leg 11, HOTO leg 1 and Origin leg 18 High frames: corrected detail, bodies and
+  equipment render; Origin's source-opaque roots-village cards remain visible.
+- Final Stockade fallback selects **Bazil Thredd**, passes default/High settings
+  with zero errors, and has a root-reviewed High screenshot. Evidence is in
+  `graphics-fallback-surface-v5-20261006`; targeted movement is in
+  `production-targeted-routes-surface-v5-20261006`.
+- Known limits include Raksha entry-camera clipping/blocked first route, the
+  disconnected Origin Falynum approach, and source-opaque root cards. Kulech's
+  transient maroon obstruction does not recur in 25 consecutive High frames;
+  exact object attribution remains unproven and no lighting defect is demonstrated.
+- Unique HOTO v2 screenshots have frozen hashes. Earlier shared v1 screenshot
+  paths were overwritten and are marked unavailable; preserved historical asset
+  identities do not imply those old image files remain. No full clears, balance
+  certification, staging, commit or external deployment are claimed.
+- Report: `docs/ROM_ENVIRONMENT_REVIEW_2026-10-06.md`; frozen evidence and guarded
+  promotion backups: `artifacts/playtest/rom-deep-review/`.
+- Final v5 entry QA completes 30/30 unique RoM dungeons in Standard/High, with
+  60 screenshots, ready assets, grounded players and zero browser errors.
+  Independent/root contact-sheet review finds no actionable new entry regression;
+  known framing limits remain and Kulech's maroon obstruction does not recur.
+  `production-entry-surface-v5-20261006` and `validation-summary.json` retain the
+  frozen identities, source-resolution audit and final movement/settings checks.
+- Local `dist` is replaced by the reviewed v5 build only after final QA. All
+  5,967 files match the tested snapshot byte for byte; the previous build is
+  preserved at `pre-surface-v5-dist-backup`. `local-build-install-proof.json`
+  records the installation, and `release-check-final.log` records final hygiene.