healer-man/docs/ROM_ENVIRONMENT_REVIEW_2026-10-06.md

# RoM dungeon texture, lighting, and route review — 2026-10-06

This continuation reviews all 30 base RoM dungeons. Luna subprocesses handled
source-material inspection, original lighting recovery, and independent runtime
review. Screenshots from an isolated asset camera establish visual coverage;
ordinary production UI and keyboard input establish entry and walking evidence.
Neither method certifies every corridor, a full clear, or combat balance.

## Texture and material changes

The all-dungeon room review covers 30 entrances, 66 boss anchors, 74 diagnostic
corridor samples, and 36 mapping-hotspot views. The corridor cameras are samples
between authored anchors, not proof of connected walking routes. Large low-rank
UV metrics require source/material attribution before applying repeat tiling:
waterfall sheets and flow effects must not be treated as stone floors.

Six dungeons had exact native `gate_waterfall` materials exported as opaque,
with the compressed image losing its alpha: Cave of the Water Dragon, Kulech
Bones Nest, Mystic Altar, Raksha Temple, Sardo Castle, and Tomb of Seven Heroes.
The source-pinned material patch covers 40 native bindings (38 normal blends,
two additive bindings). Native prelit metadata also retains its verified unlit
semantics. The JSON-only alpha patch retains identical BIN bytes. The subsequent
unlit semantics pass can repack BIN layout; a final decoded-layout proof verifies
identical positions, normals, indices, every UV stream, morph targets, node
hierarchy/transforms, and exact PNG image bytes across all six candidates.
Optimization disables simplification and automatic texture resizing; the KTX2
decode proof demonstrates retained partial alpha, not lossless codec equality.
Mystic Altar's reviewed support-floor texture remains part of the canonical
input and its metadata is preserved.

The durable converter reads native alpha parameters and exact diffuse-channel
joins. `lib/native-alpha-cache-validation.mjs` rejects old resource caches unless
source hashes, exporter/converter/alpha-policy hashes, and raw/prepared native
material reports agree. `stage-native-waterfall-alpha.mjs` produces guarded
review candidates; `promote-native-waterfall-alpha.mjs` verifies identities and
creates exclusive baseline backups before updating canonical/shipping assets
and their current hashes. Historical proofs remain historical.

The Lair of the Demon Dragon's `magic_magma_02` sheets had the same provenance
gap. All three exact source rows (2, 3, 9) are prelit cutouts, with native alpha
test value 192. They now retain MASK cutoff 192/255 and unlit semantics, replacing
the opaque black slab with the source orange flow strips. The 512×128 UASTC
image retains alpha 0–255; source and decoded image each have 37,654 pixels
passing `>192`. Mean alpha error is 0.000244, so codec equality is not claimed.
The final canonical proof preserves all geometry, UVs, hierarchy/transforms, and
all 17 exact source PNG images. These are native flow-effect sheets, not stone
floor atlas mappings, so no repeat tiling was guessed. Its stage, response-body
SHA checks, first-hit material proof and screenshots live in
`artifacts/playtest/rom-room-coverage/lair-magma-alpha-stage/`.

The six waterfall acceptance pairs require the exact target texture to be the
nearest ray hit in both shipping baseline and compressed candidate. The served
candidate identity and native BLEND provenance are verified. Four-metre Kulech
and Tomb pairs supply additional surroundings; Kulech's additive effect is very
bright and is not credited as a complete scene-quality pass. Early bounding-box
views and the first misrouted comparison are diagnostic failures, superseded by
`gate-waterfall-triangle-pairs-report.json` and
`gate-waterfall-context-pairs-report.json`. Synchronous camera scans explicitly
refresh the camera matrix before raycasting; otherwise they can inspect the
previous pose between rendered frames.

The final Barren policy replaces whole-atlas repeats with four exact 256×256
native floor islands. Only triangle corners contained in one reviewed island
receive the ×4 presentation UV1; source UV0, geometry and original atlas remain.
Its walls use a fifth exact 256×256 crop from the native central wall region at
[256,0], sampled at two metres with mirrored edges and rotated phase blending.
Wall base contrast is 0.25 around the source atlas mean, with detail strength
0.85; floor detail strength is zero. This explicit browser treatment reduces
oversized smearing and obvious tile repetition without generating replacement
pixels. The original 47,084 triangles and eight native PNG images are retained.
Its exact crop RGBA proofs are separate from lossy compressed-texture metrics.

HOTO keeps its native non-repeating cliff artwork and UV0. Its revised v2 policy
samples the original full stone/moss tiles directly at two metres on floor faces
of materials 26 and 18. Those floors use base strength 1 and modulation strength
0; material 18 samples its own moss tile. Cliff materials 96 and 53 retain their
base artwork, with floor/wall detail strength 0.85, wall contrast 0.6 around each
source atlas mean, and rotated detail variation 1. Material 26 retains wall
strength 0; material 18 retains 0.45. These settings correct presentation scale,
not native texture resolution. The v1-to-v2 patch changes only approved material
metadata; all non-JSON chunks remain byte-identical in canonical and optimized
GLBs. All 189,855 source triangles and 17 native PNG images are retained. Source-detail
carriers require joining/pruning protections during optimization. HOTO adds no
cropped images; its decoded source-layout proofs are separate from lossy
compressed-texture metrics. The fixed-camera compressed comparisons use verified served hashes and
report no browser errors. Evidence is under
`artifacts/playtest/rom-deep-review/uv-stretch-review/`.
The v2 continuation is promoted through `promote-hoto-surface-v2.mjs`, with a
new exclusive recovery backup and `hoto-surface-v2-promotion-proof.json`.
Shipping SHA-256 is `b80b60cba6e5a0eb3e13bd7e048f9408f21bf870c0c81fc8de5c195d3a3eaee3`.
The unique v2 screenshots have pinned hashes in `hoto-surface-v2-visual-proof.json`.
The earlier shared HOTO v1 final screenshot names were overwritten during the
v2 comparison and are explicitly marked unavailable in its stage/promotion
reports. Their historical asset hashes remain recorded; those image paths must
not be treated as preserved v1 captures. Earlier diagnostic captures remain
separate evidence. The current unique v2 proof supersedes that comparison.

## Source lighting and shadows

The native zone exporter recovers environment settings and oriented zone
triggers from all 30 source WDBs. Generated data retains source hashes, native
ambient/primary/two secondary colors and directions, fog, time/weather groups,
and priority/parent relationships. The resolver converts runtime positions back
into source space and follows native trigger selection and setting inheritance.
Time/weather groups default to 0/0; this does not implement a live weather clock.

`DungeonLighting.tsx` supplies these settings to production RoM environments.
Rune colors are linear byte/255 values. Rune's unoccluded SH constant band is
supplied as ambient fill (4/17 of the weighted directional colors), with the
complementary 13/17 in Three.js directionals. This preserves readability in
roofed rooms without adding a guessed white light. Three.js angular diffuse and
PBR specular remain an adaptation; the original separately authored specular
path and full SH shader are not reproduced. Standard's source-primary shadow
follows the player's footing; High follows the camera. Missing native zones use
the reviewed presentation profile. WoW uses its existing profile lighting path.

The final adapted-source entrance check covers all 30 dungeons: 29 native-zone
selections and Grafu Castle's entrance profile fallback, with zero browser
errors. The broader source-lighting camera review covers 96 entrance/boss
anchors, 95 native-zone selections and the same one fallback. Four same-camera
profile comparisons establish that SH fill keeps Forsaken Keeper 3 and Ystra's
Raichika room readable; source-dark Sardo and Kulech remain dark interiors.
Barren's pale source lighting retains rock/sand detail without clipped whites.

## Origin water

All 126 source water layers retain their native GUID joins, visibility,
vertex RGBA, source tile coordinates and 30-frame normal sequence. Metadata now
also retains the first source vertex's split-plane height and exact serialized
full-model enable flag. Raw effect value 4 keeps its original value and unknown
enum name; Rune's non-DIRECT full-enable branch establishes its rendering policy.

The browser reflects the camera across that native plane, clips below it, and
captures the pre-water scene for refraction. It uses Rune's shared projected UV,
0.025 bump perturbation, and 0.02037/0.97963 Fresnel coefficients. Per tagged
asset root, four active 512-pixel target pairs share at most ten two-pass
captures per second. Oldest-request scheduling avoids render-order starvation;
inactive pairs release their slots when a later plane needs them. Materials
dispose shared targets once. Additional active planes and below-plane views use
the lit animated fallback. Reflect-class filtering and the native water SH
shader remain unsupported. Water stays visual-only.

Same-camera final water captures record visible glancing reflections, zero
browser errors, and approximately 60 FPS. Metadata-only promotion preserves
both canonical and optimized BIN hashes. Original-source water colors remain
unchanged. Evidence: `artifacts/playtest/rom-deep-review/water-screen-effects/`.

## Walking evidence and remaining route issues

Before this lighting/material continuation, ordinary UI walking covered all 30
current source environments: 489 attempted legs, 487 reached, 882/882 grounded
samples, 1,010.312 m travelled, and zero browser errors. Twenty-eight planned
walks completed. Raksha's first forward probe blocked while grounded; Dalanis's
second waypoint exhausted the movement time budget while grounded. These are
retained failures, not successful clears. See
`artifacts/playtest/rom-deep-review/ordinary-walking-baseline/results.json`.

The complete Origin source review still finds no supported entry-to-Falynum
connection. Full 256-chunk terrain and current options-aware collision/navigation
leave the Falynum platform disconnected. Its visible ROS platform is 7.23 m
above the ground below, and the native path component is separate. The referenced
boss Lua teleport is a combat catch-up action after acquiring an attack target,
not an approach transport. No bridge or approach teleport was inferred. The
entry-to-Life connection remains in the current Recast navigation and ordinary
walking proof. Detailed verdict:
`artifacts/playtest/origin-route-diagnostic/origin-falynum-source-connectivity-verdict.md`.

## Evidence and validation

- Room cameras: `artifacts/playtest/rom-room-coverage/`.
- Compressed alpha candidates and guarded promotion:
  `artifacts/playtest/rom-room-coverage/native-alpha-stage/`.
- Adapted lighting: `artifacts/playtest/rom-source-lighting-adapted-entry-qa/`;
  comparisons: `artifacts/playtest/rom-source-lighting-profile-compare/`.
- Native lighting resolver, linear runtime colors, source shadow pose,
  terrain/water materials, water resource lifecycle/scheduling, native alpha
  cache/staging/converter semantics, source UV exceptions, and optimizer policy
  have focused passing tests. TypeScript checks pass.

The pre-v2 reviewed build is `artifacts/playtest/rom-deep-review/build-reviewed-20261006/`.
Its frozen manifest is `reviewed-content-manifest-20261006.json`, version
`content-fcd09f7c0f8f10be`, SHA-256
`f1a7ab2dc9a3d05513f3601ed5142011f71aa58155a7ec1d34bdc42e4f9309af`.
TypeScript, Vite production build, static precompression and release-bundle
verification pass: 1,328 modules, 126 chunks and 5,569 bootstrap objects.
Its all-30 quality audit matches 1,620/1,620 images to source resolution,
including five exact pixel crops, with zero downsizing, 2048-cap hits or issues.
This audit skips general pixel decoding; targeted codec/alpha proofs described
above cover the changed material families. All 162 dungeon GLBs validate with
zero actionable errors/warnings. KTX2 audit covers 106 visual assets and 12,897
compressed images with zero fallbacks; the existing performance budgets have
zero regressions. Focused runtime tests pass 83/83 and focused pipeline tests
pass 17/17. No readiness promotion, full balance sweep, commit, or external
deployment is part of this review.

The refreshed v5 build is `artifacts/playtest/rom-deep-review/build-surface-v5-20261006/`,
with index SHA-256 `ab46b9c47a0ba27b0b1c2e0a2c65ff02c613b7b87c54d37fe5ec9a7a41707e27`.
Its frozen manifest is `surface-v5-content-manifest-20261006.json`, version
`content-d1ec6f1f858f21c1`, SHA-256
`e102c30cc47a4130fff88809ce53453929e86e6f766b0afbbae8a8ad4a52fd7c`.
TypeScript, production build, 129-source static precompression (258 outputs) and
release-bundle verification pass. It retains 126 chunks and 5,569 bootstrap
objects. The refreshed all-30 quality audit again matches 1,620/1,620 textures,
including five verified crops, with zero downsizing, cap hits or issues. All 162
dungeon GLBs validate with zero errors/warnings. Focused checks pass 83 runtime
and 17 pipeline tests. See `environment-quality-surface-v5-20261006.{json,md}`.

The v5 ordinary walking checks reached 48/48 legs across Barren (27), HOTO (2)
and Origin (19), with 89/89 grounded samples, 101.826 m travelled and zero errors.
HOTO remains a six-metre probe; Origin follows Life. Root inspected Barren leg 11,
HOTO leg 1 and Origin leg 18 High frames: corrected detail is present, bodies and
equipment render, and Origin's documented opaque root cards remain. Evidence is
in `production-targeted-routes-surface-v5-20261006/`. Stockade's v5 fallback
selected **Bazil Thredd**, passed default/High settings and recorded zero browser
errors; root inspected the High frame. Its evidence is in
`graphics-fallback-surface-v5-20261006/`.

The final v5 entry pass reaches all 30/30 registered dungeons with ready assets
and grounded players in both Standard and High (150% render scale, 16× filtering,
authored ambient). It records 60 screenshots and zero browser errors. Independent
and root contact-sheet review found no actionable new entry regression; the
documented Raksha/source-panel framing limits remain, and Kulech's maroon
obstruction did not recur. See `production-entry-surface-v5-20261006/`.
`validation-summary.json` independently checks the frozen build/manifest, all 30
unique dungeon IDs, settings, movement outcomes and the three changed visual
assets against both shipping files and immutable content objects.
The local `dist` now contains that reviewed build: all 5,967 files match byte for
byte. The previous build is preserved at `pre-surface-v5-dist-backup/`; installation
proof is `local-build-install-proof.json`. This is a local installation, not an
external deployment. Final release hygiene is recorded after note updates in
`release-check-final.log`.

The pre-v5 intermediate frozen production pass rendered all 30 entrances through the
normal UI, with ready assets and grounded players in both Standard and High,
60 screenshots, and zero browser errors. The eight changed material/water
environments also received ordinary keyboard walking: 150/151 legs reached,
277/277 grounded samples, 325.037 m travelled, and zero browser errors. Raksha's
blocked first leg is retained. Stockade's graphics fallback selected Hamhock and
passed default/High rendering checks. Evidence is under
`artifacts/playtest/rom-deep-review/production-entry-final/`,
`production-walking-final/`, and `graphics-fallback-final/`.

The pre-v2 targeted routes use the reviewed build and frozen manifest above:
Barren reached 27/27 legs, HOTO 2/2 and Origin 19/19. All 95 sampled positions
were grounded, with 103.622 m travelled and zero browser errors. HOTO remains
only a six-metre collision probe; Origin targets the connected Life route,
not Falynum. Standard/High walking frames are in
`production-targeted-routes-reviewed-20261006/`. That Stockade fallback
selected **Kam Deepfury** and passed default/High settings, with zero errors;
its evidence is in `graphics-fallback-reviewed-20261006/`.

The pre-v2 all-30 entry pass also reached ready, grounded entries in Standard
and High with 60 screenshots and zero browser errors. Visual review flagged
remaining HOTO scale issues, addressed by the v2 policy above. Barren retains
broad source-painted striations despite the corrected fine detail. Raksha's
entry camera clips inside architecture and its first route leg remains blocked;
this is not a complete scene-quality acceptance. Wider source-anchor views
provide separate material/lighting coverage. Kulech's High entry frame contained
a maroon obstruction during combat. It did not recur in 25 consecutive High
frames in an ordinary-UI follow-up. The original Standard/High pair compares
changing combat states, so exact object attribution remains unproven; no lighting
defect was demonstrated. Read-only rays in the clear frames identify the static
source wall/stone behind it. See `kulech-obstruction/findings.md` and `frames/`.

Those intermediate production screenshots found further issues even though the
historical pre-crop audit's 1,615 matched textures retained their dimensions. Resolution checks do
not establish good surface mapping. Barren's whole-atlas repeat mixes floor and
wall regions; restoring UV0 removes the added bands but leaves oversized detail.
The promoted atlas-safe floor/wall correction and HOTO source-detail policies
above supersede those intermediate mappings.
Origin's giant-tree family also retained alpha pixels but lacked native MASK on
six canonical materials merged into one optimized material. Its source-pinned
JSON-only correction is now promoted, preserving every non-JSON GLB chunk and
all terrain/water settings. Seven freshly exported ROS resources prove standard
lighting and MASK cutoff 192/255; cache/converter smoke passes all seven. The
same-camera 3.5 m leaf close-up has matching served hashes, exact first hits,
unchanged lighting and zero errors. A separate leg-18 diagnostic context uses
the recorded walking position, but is not itself ordinary input evidence.
The separate `elves_vine` card also lost cutout semantics and compressed alpha.
Fresh exports independently prove that `elves_vine.ros` material 1 and
`act_green_tower.ros` material 3 use the same DDS with standard lighting and
MASK cutoff 170/255. Both canonical bindings and their optimized merged material
now retain that policy. The replacement 256×256 UASTC image retains nine mip
levels and all 10,407 source pixels passing the cutoff, with zero threshold-mask
mismatches; alpha bytes are not claimed lossless. The canonical BIN is unchanged,
and the optimized patch preserves all non-target buffer views, mesh streams,
terrain and water metadata. Same-camera served-hash proof removes the black
vine panel with zero browser errors. Its exclusive backup and proof live under
`artifacts/playtest/rom-deep-review/origin-root-card-attribution/`.
The residual leg-18 card is `Object01_1_mesh_0_10`, native material 1 from
`gaint_tree_a.ros`, using `elves_gaint_roots_village`. Fresh source flags request
alpha test 192, but all eight native DXT1 mips, the prepared PNG, and all ten
shipping KTX mips are fully opaque. This is not a lost-alpha regression. The
exporter supplies neither vertex-color/alpha nor face-culling channels, so the
complete native behavior remains unverified. No black-keyed mask or forced
cutout was invented. The card remains a visual limitation, documented in
`origin-root-card-attribution/residual-roots-village/roots-village-alpha-proof.json`.
Evidence and exclusive promotion backup:
`artifacts/playtest/rom-deep-review/origin-tree-alpha/`.

Mystic's crimson entry panel is the source-authored opaque
`cowcowcowcow01_164_mesh_0`; the source atlas is entirely opaque. Sardo's close
column is the authored `channel_pillar_01_52_mesh_0`, likewise opaque. Sardo's
first ordinary walking leg opens the room; Mystic's panel still occupies much
of some walking frames, so those frames do not establish whole-room lighting.
The source-anchor camera review supplies wider material/lighting coverage.
Exact source identities and framing evidence are in
`production-entry-final/entry-framing-attribution.md` and its JSON companion.