description: "Per-game audit of Fallout 4 compatibility — BA2, half-float verts, BGSM materials, M49 precombines (CSG)" argument-hint: "--focus <dimensions>"

Fallout 4 Compatibility Audit

Deep audit of ByroRedux readiness for Fallout 4 content.

Architecture: Orchestrator. Each dimension runs as a Task agent (max 3 concurrent).

See .claude/commands/_audit-common.md for project layout, game data locations, key reference docs, methodology, deduplication rules, and finding format. See .claude/commands/_audit-severity.md for severity. This file only carries FO4-specific context — do not duplicate the common map here.

What's Landed (audit as regression guards, not blockers)

FO4 is one of the more complete game paths. Treat each item below as shipped — the audit job is to confirm it still holds, not to re-propose it.

NIF — BSVER 130 + Next-Gen patch range parses. Half-float vertices, BSTriShape, BSGeometry, inline tangents, BSSubIndexTriShape.
BA2 — exhaustive version match {1, 2, 3, 7, 8} (GNRL + DX10); unknown majors bail at open().
BGSM/BGEM materials — full external-material parser crate crates/bgsm/, merged into the mesh before NIFAL translate.
ESM architecture records — SCOL / MOVS / PKIN / TXST parsed; SCOL/PKIN placements expand into individual statics.
M49 precombined geometry (CSG) — closed Session 45, 2026-06-02 (#1351 / #1188 Stage A). .csg reader, NIF precombine decode, cell-loader spawn, per-tier LOD selection, owning-REFR texture slot routing all shipped. ROADMAP language describing M49 as "blocked / no spec" is stale — the format was cracked from first principles, spec written, pipeline landed.
FO4 metalness from spec chromaticity (#1476) — legacy spec-glossiness BGSMs derive metalness from spec-color saturation, not luminance (luminance made all vanilla concrete read chrome).

Still open (legitimate forward scope): _precomb.nif collision, .uvd occlusion volumes, MOVS physics runtime, FaceGen NIF truncation tail, deeper cell coverage (LIGH power state / CONT leveled items / NPC_ face morph), Havok behavior-graph driving.

Game Context

| Aspect | State | |--------------|-------| | NIF format | BSVER 130 + Next-Gen patches; FALLOUT4..FO4_DLC_UPPER (130..=139) carries DLC trailing fields (crates/nif/src/version.rs) | | BA2 format | Exhaustive match over {1, 2, 3, 7, 8} in crates/bsa/src/ba2.rs (consts BA2_V_FO4=1, BA2_V_FO4_NEXT_GEN_TEX=7, BA2_V_FO4_NEXT_GEN_MESH=8); GNRL + DX10 | | ESM records | SCOL / MOVS / PKIN / TXST in crates/plugin/src/esm/records/; SCOL/PKIN expand in byroredux/src/cell_loader/refr.rs | | Precombines | M49 CSG pipeline landed — crates/bsa/src/csg.rs → crates/nif/src/import/precombine.rs → byroredux/src/cell_loader/precombined.rs | | Parse rate | 100.00% clean on both vanilla mesh archives per #1457 (parse_rate_fo4_all_meshes, 2026-06-14); the FaceGen truncation tail the 2026-06-02 ROADMAP snapshot recorded is gone. ROADMAP compat-matrix now also reads 100.00% (159 866/159 866, #1593) — still re-run the harness before citing a parse rate. | | Rendering | Interior cells render end-to-end (MedTekResearch01 bench, ~21k entities). BGSM/BGEM merged; precombined entities spawned interior + exterior. | | Reference | /mnt/data/SteamLibrary/steamapps/common/Fallout 4/Data/ | | Bench | --cell MedTekResearch01 (see ROADMAP for the full --bsa/--textures-ba2/--materials-ba2 invocation) |

Known FO4 Specifics

Half-float vertices — VF_FULL_PRECISION controls f32 vs u16 binary16 positions/normals; FO4 defaults to half.
FO4 shader flags — BSLightingShaderProperty flags are a u32 pair (shader_flags_1 + shader_flags_2), not the single mask of FO3/FNV/Skyrim.
DLC trailing fields — subsurface, rimlight, backlight, fresnel, wetness in the FALLOUT4..FO4_DLC_UPPER BSVER band.
External materials — PBR-ish params live in .bgsm/.bgem files; the NIF carries only the path. The block parser returns a material-reference stub when BSVER ≥ 155 and Name is a BGSM/BGEM path; crates/bgsm/ parses the file.
Inline per-vertex tangents (#795 / #796) — when VF_TANGENTS | VF_NORMALS are both set, FO4+ BSTriShape ships tangents inline in the packed-vertex blob, NOT via a separate NiBinaryExtraData (the Skyrim/FO3/FNV path, #786). Consolidating the two into one is a regression of #795/#796.
Architecture records — SCOL (prefab collections), MOVS (movable statics), PKIN (packins), TXST (texture sets w/ DODT decal-data + DNAM flags).

Parameters (from $ARGUMENTS)

--focus <dimensions>: Comma-separated dimension numbers (e.g., 1,3). Default: all 9.

Phase 1: Setup

Parse $ARGUMENTS.
mkdir -p /tmp/audit/fo4.
Dedup baseline: gh issue list --repo matiaszanolli/ByroRedux --limit 200 --json number,title,state,labels > /tmp/audit/issues.json.
Confirm Fallout 4/Data/ exists; if not, note which dimensions lose real-data validation.

Phase 2: Launch Dimension Agents (parallel)

Dimensions are ordered by FO4 risk: the precombine pipeline, BGSM material translation, and the BA2 reader are the hot, recently-churned areas.

Dimension 1: M49 Precombined Geometry — CSG + Decode + Spawn (highest FO4 risk)

Subagent: general-purpose Entry points: crates/bsa/src/csg.rs (CsgArchive::open, read_psg), crates/nif/src/import/precombine.rs (decode_shared_geom_object, psg_vertex_stride, PrecombineGeometry::into_imported_mesh, collect_precombine_geom_refs, PrecombineMaterial::apply), byroredux/src/cell_loader/precombined.rs (spawn_precombined_meshes), byroredux/src/cell_loader/load.rs (the pc_spawned / absorbed_refs gate), crates/plugin/src/esm/cell/wrld.rs + crates/plugin/src/esm/cell/mod.rs (precombined_mesh_hashes). Spec: docs/engine/fo4-csg-format.md. Checklist:

Vertex stride — psg_vertex_stride(vertex_desc) must match the BSPackedGeomObject packed layout; an off-by-N stride silently shifts every vertex. Cross-check against the spec doc.
Y-up convert — into_imported_mesh applies the Z-up→Y-up + per-instance transform. Confirm the instance transform composes (no double-apply with cell_origin).
CSG resolve — spawn_precombined_meshes opens the companion <Plugin> - Geometry.csg once per cell; missing CSG must fall back to per-REFR (return 0 spawns) rather than panic.
Owning-plugin resolution (#1592→#1590, d6bf8437) — the CSG path AND the _oc.nif precombine-object path key off the cell's owning plugin (form-id high byte), NOT the last-loaded --esm (precombined.rs::open_geometry_csg is fed the owning path). A DLC / multi-master cell must open <owning-plugin> - Geometry.csg, else every DLC-owned precombine silently falls back to per-REFR (or reads the wrong CSG blob). Tests: the #1590 cases in precombined.rs.
REFR de-dup gate (#1188) — cell.absorbed_refs are suppressed only when pc_spawned > 0; verify the load.rs gate renders absorbed REFRs when the precombine spawns nothing (no double-draw, no holes).
LOD selection — one tier per object, not all three (the a30c088a fix); a regression here triples precombine geometry.
Texturing — precombine shapes texture through the owning REFR's shape-slot indices (2900de70); a regression renders untextured/checker precombines.
Opaque-architecture alpha-blend guard (887aae52) — on the precombine path only, spawn_precombined_meshes keeps the BGSM merge's two_sided/decal/alpha_test/texture flags but restores the pre-merge (NIF-shape) alpha-blend state. FO4 authors the "Standard" blend mode identically on transparent lab glass and opaque Institute metal architecture, so forwarding the merged has_alpha makes every precombined wall MATERIAL_KIND_GLASS → see-through, mirror-hazy walls. A regression that re-applies the merged alpha-blend on the precombine path reintroduces the transparent/reflective-wall bug.
Exterior path — cell_origin = cell_grid_to_world_yup(gx, gy) for exterior (#1221 / #1222); without it exterior precombines stack at world origin. Output: /tmp/audit/fo4/dim_1.md

Dimension 2: BGSM / BGEM Consumption + Metalness-from-Chromaticity (regression pin #1476)

Subagent: general-purpose Entry points: crates/bgsm/src/ (lib, base, bgsm, bgem, reader, template), byroredux/src/asset_provider/material.rs (merge_bgsm_into_mesh). Checklist:

Single data source — merge_bgsm_into_mesh is the FO4 material merge: albedo/diffuse tint, specular, emissive, smoothness→roughness, translucency suite (#1147), model-space-normals bit, texture-slot paths. It runs before translate_material.
Metalness from saturation, NOT luminance (#1476, 08ed03be) — for legacy spec-glossiness BGSMs (leaf.pbr == false, ~all vanilla architecture), metalness = (max-min)/max of specular_color (mult-invariant saturation): white spec [1,1,1] → 0 (concrete is dielectric), tinted spec → metallic. Any reversion to luminance (0.2126·r + …) for the non-pbr branch is the #1476 regression — it makes vanilla concrete read chrome. The luminance path is correct only for leaf.pbr == true.
smoothness→roughness applied exactly once at parse — mesh.roughness_override = Some((1.0 - leaf.smoothness).clamp(0.04, 1.0)). Audit any downstream re-application of the 1.0 - smoothness inversion.
Magic-vs-extension reconciliation (#758 footgun) — the .bgsm/.bgem/.mat arm dispatches on file content; verify it logs and does not blindly trust the extension.
Cycle-aware template resolve (#1148) — template.rs::resolve / resolve_depth track visited lowercase keys and break A→B→A self-refs (vanilla defaulttemplate_wet.bgsm); verify the cycle-break + DEPTH_LIMIT still fire.
Blend-factor pass-through, NOT a swap (#1823 fixed a regression of #1651, 27334481) — src_blend/dst_blend in BGSM/BGEM are already Gamebryo-native (bgsm_blend_to_gamebryo, a plain u8 cast); Standard=(6,7), Additive=(6,0), Multiplicative=(4,1) feed gamebryo_to_vk_blend_factor verbatim. #1651's premise ("GL-style enum inverted at 0/1") was false and its 0↔1 swap corrupted Additive (dst 0→1→ZERO, kills accumulation) and Multiplicative (dst 1→0→ONE, dest leaks through); Standard's (6,7) is a fixed point of the swap, which is why it slipped through review. Any reintroduction of a 0↔1 swap (or a function named gl_to_gamebryo_blend) is the regression. Output: /tmp/audit/fo4/dim_2.md

Dimension 3: BA2 Reader — GNRL + DX10 (Exhaustive Version Match)

Subagent: general-purpose Entry points: crates/bsa/src/ba2.rs. Checklist:

Exhaustive dispatch (#811, f480337) — version handling is a match over the supported set; unknown majors (0, 4, 5, 6, 9, …) bail at open(). Any reintroduction of a cascading if v == 1 { … } else if v == 7 { … } chain with a silent v1 fallback is the regression. Mirrors the BSA allowlist now at crates/bsa/src/archive/open.rs (the old archive.rs:165-173 comment in ba2.rs referencing the pre-split file is stale).
GNRL — extract a mesh and a script, byte-exact.
DX10 — DDS header reconstruction from width/height/dxgi_format/mip count; DXT1/DXT5/BC5/BC7 encoding; mip chunk assembly (mip0 largest → mip_last); per-archive zlib-vs-uncompressed flag.
Cross-version — BA2_V_FO4 (1) / BA2_V_FO4_NEXT_GEN_TEX (7) / BA2_V_FO4_NEXT_GEN_MESH (8) all force zlib; confirm the Next-Gen mesh/tex archives extract. Output: /tmp/audit/fo4/dim_3.md

Dimension 4: NIF BSVER 130 + Half-Float Vertices + FO4 Collision

Subagent: legacy-specialist Entry points: crates/nif/src/blocks/tri_shape/bs_tri_shape.rs, crates/nif/src/import/mesh/bs_tri_shape.rs, crates/nif/src/import/mesh/bs_geometry.rs, crates/nif/src/import/collision/mod.rs. Checklist:

VF_FULL_PRECISION resolution (default-half unless set); half-float decode matches IEEE 754 binary16 incl. denormals/NaN.
BSSubIndexTriShape segment data walked (FO4 actors lean on it); skinned bone indices/weights honor packed layout.
DLC trailing fields on BSLightingShaderProperty in the FALLOUT4..FO4_DLC_UPPER band (subsurface/rimlight/backlight/fresnel/wetness). Next-Gen patch BSVER values still dispatch.
#1242 constant rename — the BSVER bound is FO4_DLC_UPPER (= 140), not the old misnamed FO4_ENV_SCALE. Any reference to the old name anywhere under crates/nif/ or crates/plugin/ is a regression target.
FO4 collision (regression pin) — BhkMultiSphereShape and BhkConvexListShape translate to CollisionShape in collision.rs (search the two downcast_ref arms) instead of falling through to the "unsupported" log + drop. Power-armor / settlement / destructible-debris collision depends on this; either arm reverting to a drop is a regression (NIFAL "parsed then dropped" leak class). Output: /tmp/audit/fo4/dim_4.md

Dimension 5: FO4 Shader Flags & BGSM PBR Routing (Disney path)

Subagent: renderer-specialist Entry points: crates/nif/src/blocks/shader.rs (BSLightingShaderProperty), crates/nif/src/import/material/ (mod, walker, shader_data), crates/renderer/shaders/include/lighting.glsl + include/pbr.glsl (Disney lobe gates). Checklist:

u32 flag pair read as two separate fields (shader_flags_1 + shader_flags_2). FO4 flag bit positions differ from Skyrim — verify decal / alpha-test / skinned / glow / window / refraction / parallax / facegen bits read from the correct mask + bit.
Render-affecting bits consumed, not just parsed (#1592, f7fbbed5) — the walker (crates/nif/src/import/material/walker.rs) must OR Model_Space_Normals (F4SF1 bit 12) and Alpha_Test (F4SF2 bit 25) into MaterialInfo. Glow_Map (F4SF2 bit 6) is deliberately not gated here (#1733 doc-fix, corrected the #1592 comment) — glow sources unconditionally from the BSShaderTextureSet glow slot regardless of the flag bit; do not re-propose wiring it as a gap. Vanilla FO4 is BGSM-backed so this only bites inline/loose/modded NIFs, but a regression decodes an object-space normal map as tangent-space or renders an inline cutout opaque. Tests: crates/nif/src/import/material/fo4_shader_flag_tests.rs, alpha_flag_tests.rs.
BGSM Name stopcond — when the material is external, the block parser returns before reading the inline Phong trailing fields (those belong to the BGSM). ImportedMesh.material_path flows to Material.material_path for diagnostics; mesh.info <id> surfaces it when texture_path is absent.
BSShaderTextureSet slot routing (#563) — SkinTint / HairTint BSLightingShaderType variants sample different slots than the default LSP path; verify the per-type slot map is in lockstep with what triangle.frag reads.
PBR flag, not Lambert — crates/nif/src/import/material/walker.rs routes BGSM-authored material through the PBR path; the FO4-specific regression is a BGSM material falling back to Lambert (opposite of the FNV/FO3 regression). #1241 surfaces smoothness/IOR/specular_strength into MaterialInfo so the Disney lobe has data.
#1244 BSShaderPropertyBaseOnly consumer wired at import — bare-stub BSShaderProperty routes through MaterialInfo; verify FO4 fallback meshes aren't default-everything.
The per-game→canonical handoff is covered in depth by Dimension 7; cross-game canonical invariants live in /audit-nifal. Output: /tmp/audit/fo4/dim_5.md

Dimension 6: ESM Architecture Records (SCOL / MOVS / PKIN / TXST)

Subagent: general-purpose Entry points: crates/plugin/src/esm/records/ (scol.rs, movs.rs, pkin.rs, mswp.rs; TXST lives in the misc set), crates/plugin/src/esm/cell/ (mod, walkers, support, wrld + cell/tests/). Checklist:

SCOL — prefab of child-static placements w/ per-instance scale/rotation, parsed into an expandable structure. PKIN — packins (grouped bundles). MOVS — movable statics (parse-only; physics runtime not wired). TXST — texture sets referenced by NIF material paths.
TXST DODT + DNAM (#813 / #814) — decal-data sub-record + flags parsed via DecalData; pre-fix, 207/382 (DODT) + 382/382 (DNAM) vanilla TXSTs dropped authoring. Any TXST read path without DecalData is the regression. The TXST match arms live in cell/walkers.rs; an unreachable_patterns warning there suggests b"TXST" matches before the intended arm.
Category index exposure (#817) — the FO4-architecture maps must surface in categories(); a missing entry hides records from category iteration.
Re-run the real-data ESM parse harness (#819) before reporting any parse-rate finding. Output: /tmp/audit/fo4/dim_6.md

Dimension 7: NIFAL Canonical Material Translation (FO4 is PBR-canonical)

Subagent: renderer-specialist Entry points: byroredux/src/material_translate.rs (translate_material — the single boundary), crates/core/src/ecs/components/material.rs (Material, Material::resolve_pbr), byroredux/src/cell_loader.rs (pack_bgsm_material_flags, pack_effect_shader_flags). Spec: docs/engine/nifal.md. See also /audit-nifal. Checklist:

Single boundary — translate_material(mesh, paths, extra_material_flags) is the only ImportedMesh → Material site; both the cell-loader REFR-spawn path and the loose-NIF scene path must route through it, not verbatim struct literals.
BGSM PBR flag routing — effect_shader_flags ORs pack_effect_shader_flags + pack_bgsm_material_flags(mesh) + caller extra_material_flags. Verify BGSM meshes get BGSM_PBR (and where authored BGSM_TRANSLUCENCY / BGSM_MODEL_SPACE_NORMALS) → shader-side MAT_FLAG_PBR_BSDF. Tests: pack_bgsm_material_flags_tests in cell_loader.rs.
Resolve-once contract — Material.metalness / Material.roughness are plain f32 (not Option + per-draw classify), seeded from mesh.metalness_override / mesh.roughness_override (or f32::NAN for legacy inline-shader content), then resolve_pbr() fills NaN from the classifier and clamps (metalness 0..1, roughness 0.04..1). resolve_pbr is idempotent (tests resolve_pbr_is_idempotent, resolve_pbr_preserves_upstream_translator_values, resolve_pbr_fills_only_missing_slot in material.rs). Any consumer re-deriving roughness or re-running a classifier at draw-time is the regression.
Glass is classified after resolve_pbr (forced glass roughness wins) via helpers::classify_glass_into_material. Output: /tmp/audit/fo4/dim_7.md

Dimension 8: FO4 Cell Load End-to-End (SCOL/PKIN Expansion)

Subagent: general-purpose Entry points: byroredux/src/cell_loader/refr.rs (expand_scol_placements, expand_pkin_placements), byroredux/src/cell_loader/references/mod.rs (wires the expanders), byroredux/src/cell_loader/exterior.rs (Phase-3a). Checklist:

expand_scol_placements / expand_pkin_placements turn a prefab/packin into per-instance synthetic REFRs with composed transform (parent × child), recursion bounded by MAX_PKIN_DEPTH = 4 (shared, #1180 / #1182; vanilla has zero nesting, the cap guards modded cross-recursion). references.rs fires the first matching expander and composes placements.
Tests: cell_loader/scol_expansion_tests.rs, cell_loader/pkin_expansion_tests.rs.
Precombine spawn (Dim 1) and SCOL/PKIN expansion are independent carriers of the same cell — confirm no double-draw between an expanded SCOL static and a precombine that subsumed it (the absorbed_refs gate in Dim 1 is the de-dup mechanism).
BSConnectPoint attach-graph consumer (#1594, c16600a5) — the import lifts BSConnectPoint::Parents/::Children into AttachPoints / ChildAttachConnections ECS components (crates/core/src/ecs/components/attach_points.rs); the cell-load spawn path (cell_loader/spawn.rs + references.rs, cached in nif_import_registry.rs) must materialize them onto spawned entities. Pre-fix the chain dead-ended at the import boundary and modular weapons / power-armor frames spawned as base-receiver only. Regression target: a spawn path that imports the connect graph but attaches no AttachPoints component. Tests: cell_loader/attach_points_spawn_tests.rs. Output: /tmp/audit/fo4/dim_8.md

Dimension 9: Real-Data Validation + Forward Scope

Subagent: general-purpose Entry points: crates/nif/tests/parse_real_nifs.rs (parse_rate_fallout_4, parse_rate_fo4_all_meshes), crates/nif/examples/nif_stats.rs, ROADMAP.md, docs/feature-matrix.md. Checklist:

Re-run BYROREDUX_FO4_DATA=… cargo test -p byroredux-nif --test parse_real_nifs -- --ignored fo4 (parse_rate_fo4_all_meshes covers both Fallout4 - Meshes.ba2 + Fallout4 - MeshesExtra.ba2) — both archives parse 100.00% clean per #1457. Report the measured number; do not cite the stale ROADMAP figure without re-running.
Trace import_nif_scene on: a settlement workshop item, a creature (deathclaw / super mutant), a power-armor frame (heavy skinning + BSConnectPoint), a modular weapon (receiver/barrel/stock via BSConnectPoint). For each: mesh count, material_path (BGSM ref or null), skinned vs rigid, connect-point extra-data presence.
Forward scope (do NOT re-file as blockers): _precomb.nif collision, .uvd occlusion volumes, MOVS physics runtime, FaceGen NIF truncation tail, deeper cell coverage (LIGH power state / CONT leveled items / NPC_ face morph), BSBehaviorGraphExtraData parse-only (verify nothing pretends to drive it). The BGSM parser, SCOL/PKIN expansion, and the M49 CSG pipeline are all shipped — never list them as pending. Output: /tmp/audit/fo4/dim_9.md

Phase 3: Merge

Read all /tmp/audit/fo4/dim_*.md.
Combine into docs/audits/AUDIT_FO4_<TODAY>.md:
- Executive Summary — NIF + BA2 + BGSM parser + SCOL/PKIN expansion + M49 CSG precombines all landed; parse rate per the latest parse_rate_fo4_all_meshes run. Pending: precombine collision / .uvd volumes, MOVS physics, deeper cell coverage (LIGH/CONT/NPC_).
- Dimension Findings — grouped by severity per dimension.
- BGSM Consumption Table — BGSM field × merged-into-ImportedMesh-by-merge_bgsm_into_mesh / surfaced-on-canonical-Material (post-NIFAL).
- Forward Scope Chain — precombine collision / .uvd → MOVS physics → deeper REFR/LIGH/CONT/NPC_ coverage. Do NOT list the CSG reader, BGSM parser, or SCOL expansion as pending.
Remove cross-dimension duplicates.

Suggest: /audit-publish docs/audits/AUDIT_FO4_<TODAY>.md

Audit de compatibilité Fallout 4

Recommandé pour