Starfield Compatibility Audit

Regression-and-depth audit of ByroRedux's Starfield support — a first-class GameKind with NIF + BA2 v2/v3, CDB materials and a walkable Cydonia interior — not a gap inventory.

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

Read .claude/commands/_audit-common.md (layout, game data, dedup, finding format, SF Material / SF Smoke entries) and .claude/commands/_audit-severity.md (NIFAL rows gate translate_material at HIGH minimum).

Scope: Starfield's data through the shared mechanisms; mechanism defects go to the owner: BA2 / CDB reader discipline → /audit-parsers; NIF block parsing → /audit-nif; ESM walker → /audit-esm; canonical-material invariants → /audit-nifal; Starfield HUD (M48.8; the hudmenu.swf PlaceObject3 dialect shim shipped under #4470, profile + catalog not built) → /audit-ui; NPC stat model (#4453) → /audit-character.

Status authority: ROADMAP.md compat row (100.00% over 120,543 NIFs as of 2026-09-24, #4440 — still measure, never transcribe), docs/feature-matrix.md, docs/engine/starfield-esm-roadmap.md, docs/engine/starfield-esm-phase0-baseline.md, docs/audits/SF_CDB_PHASE2_SPIKE_2026-08-29.md.

Game Context

Aspect	State
NIF	BSVER 155 (FO76 baseline) → Starfield retail extensions; mesh path is BSGeometry (inline data or external geometries\<X>.mesh), never BSTriShape
BA2	v2 (zlib, 8-byte header extension) + v3 (12-byte extension with compression_method: 0 zlib, 3 LZ4 block)
Materials	CDB (crates/sfmaterial/) from materials\materialsbeta.cdb via --materials-ba2. Vanilla ships zero .bgsm/.bgem files (283 of 69,170 sampled meshes still name one → guaranteed resolver miss → CDB fallback) and no .mat sidecars (four Creation archives carry 20 loose JSON .mat files; the resolver asset_provider/material/loose_mat.rs (#4277) was re-derived against their real layout in faf09c6e6, #5396, and pinned with verbatim excerpts) — the CDB is the only vanilla material source. Phase 2 (#3398, 224a19372 / 18fce7e43 / 978d25c19 / 9bb6b7dce) translates colour/normal/emissive roles + alpha/glass/blend/emissive-settings scalars from a streaming MaterialIndex; authored metal/rough scalars, slot-6 _height (#5381) and five single-channel kinds are still parked
Cell	Walkable Cydonia interior; no runtime baseline; #3540 (frame-0 stall) fixed by plan_static_blas_restore, real-device re-run unconfirmed

Parameters / Setup

--focus <dimensions> (default all 6). Setup: parse $ARGUMENTS; mkdir -p /tmp/audit/starfield; gh issue list --repo matiaszanolli/ByroRedux --limit 200 --json number,title,state,labels > /tmp/audit/issues.json; confirm Starfield/Data/ exists (else note dimensions losing real-data validation). Read docs/audits/AUDIT_STARFIELD_2026-09-16.md first (its premise corrections), then the latest delta report (AUDIT_STARFIELD_2026-10-09.md at time of writing).

Dimensions (ordered by Starfield-specific risk)

Dimension 1: BA2 v2/v3 + Corpus Validation

Subagent: general-purpose
Paths: crates/bsa/src/ba2.rs, crates/nif/tests/parse_real_nifs.rs, crates/nif/examples/nif_stats.rs
First step: BYROREDUX_STARFIELD_DATA=… cargo test -p byroredux-nif --test parse_real_nifs parse_rate_starfield_all_meshes -- --ignored (all 13 mesh archives; parse_rate_starfield = Meshes01 only)
Guards (ba2.rs): decompress_chunk_lz4_*, v3_unknown_compression_method_rejected, extract_dx10_concatenates_a_raw_chunk_and_an_lz4_chunk_in_one_record, the extract_dx10_*_short_*_chunk* pair (#4662), dx10_open_rejects_a_non_24_chunk_hdr_len_as_invalid_data + dx10_open_tolerates_non_monotonic_start_mip_without_panicking (#5008 — both were file-controlled debug_asserts that panicked debug builds).
Checklist:

• v3: compression_method 0 → zlib, 3 → LZ4 block, anything else a hard InvalidData; LZ4 gets unpacked_size as its output bound. GNRL and DX10 both reach decompress_chunk; the per-chunk selector is packed_size == 0 = raw (v3 DX10 mips mix raw and LZ4 chunks in one texture; the sentinel is unambiguous). A DX10 non-final chunk that decodes short is rejected (it would shift every later mip); only the final chunk is lenient (#4662).
• Corpus — 100.00% clean over all 13 archives (120,543 NIFs, 0 truncated / recovered / NiUnknown; measured 2026-09-16, re-confirmed 2026-09-24 for #4440, which brought ROADMAP / game-compatibility.md / nif-parser.md in line). Confirm it stays at 0 and the texture archives extract. BSWeakReferenceNode still captures an undecoded remainder into starfield_tail (#3524's byte-audit was never done): the 0 is recovery, not decode — growth of that tail is the signal.
• Trace a clutter item, hull, body, weapon and landscape mesh through import_nif_scene; new NiUnknown = a block introduced since the FO76 baseline.
  Output: /tmp/audit/starfield/dim_1.md

Dimension 2: BSGeometry Mesh Extraction (Starfield's actual mesh path)

Subagent: legacy-specialist
Paths: crates/nif/src/import/mesh/{bs_geometry,skeleton}.rs, crates/nif/src/blocks/bs_geometry.rs, byroredux/src/asset_provider/archive.rs (normalize_mesh_path)
First step: cargo test -p byroredux-nif bs_geometry + git log --since=<last report> -- crates/nif/src/import/mesh/bs_geometry.rs crates/nif/src/blocks/bs_geometry.rs
Guards: bs_geometry_*_tests.rs (sentinel slot, resolve log, hint mismatch, skin, tangent, bounding sphere #4394); normalize_mesh_path_* in asset_provider/tests/material_path.rs.
Checklist (what the guards cannot see):

• Stage A inline (has_internal_geom_data) vs Stage B external .mesh: the canonical geometries\<X>.mesh path must NOT get a meshes\ prefix (#1292; Cydonia spawn rate collapses); iterate every LOD slot, not meshes.first() (#1209); both stages must skip a scale<=0 sentinel slot (empty vertices/triangles) even when it parses Ok/matches Internal first (#1828/#1829) — accepting it silently drops the whole BSGeometry.
• Trailer EOF — BSGeometryMeshData::parse treats EOF at the post-LOD meshlet/cull trailer as "no trailer" (facegen .mesh bodies end at the LOD array; #3777) but a body truncated mid-trailer still errors (remaining() == 0 gate). Invisible to .nif parse-rate gates (Stage B's per-slot Err arm only debug!-logs; pre-fix it zeroed all 1,282 FaceMeshes heads). The gate is explicit per path (#4269): MeshTrailer::OptionalAtEnd for standalone .mesh bodies, MeshTrailer::Required for inline Stage A bodies, where remaining() counts the rest of the NIF (guards an_inline_body_reads_its_trailer_and_stops_after_it / an_inline_body_missing_its_trailer_is_an_error). Regression guard (#4268, fixed 0abc86c8b): convert_bs_geometry_skin_weights (skin.rs) declines the whole weight set when any per-vertex bone index is >= bone_refs.len() (never a clamp; bs_geometry_skin_weights_with_out_of_range_bone_index_decline, ..._last_bone_index_is_in_range).
• Skin chain (#1203) — bone_refs are NULL on 73% of skin refs, so solve_bone_names_with_offset (skeleton.rs) fits bind-pose offsets against an externally resolved skeleton: a name is accepted only on a unique full match (MIN_BONES_TO_SOLVE = 8; offset memoised per skeleton); every decline falls back to Bone{i} (never worse). Zero wrong over 9,057 ground-truth bones; coverage partial by design (425/908 clothes skins, 8,288/22,663 bones at 2026-08-30, solver unchanged since — re-measure before citing).
  Output: /tmp/audit/starfield/dim_2.md

Dimension 3: CDB Material Database

Subagent: renderer-specialist
Paths: crates/sfmaterial/src/, byroredux/src/asset_provider/material/{cdb,loose_mat,merge,provider}.rs, byroredux/src/asset_provider/tests/starfield_mat.rs, crates/sfmaterial/examples/cdb_key_hash_probe.rs
First step: git log --since=<last report> --format='%h %cs %s' -- crates/sfmaterial byroredux/src/asset_provider/material/cdb.rs
Guards: starfield_mat.rs — discovered_cdbs_accumulate_in_load_order, mat_path_merges_cdb_authored_textures_when_indexed + mat_path_lookup_miss_keeps_presence_only_and_no_textures, cdb_hit_stamps_neutral_scalars_instead_of_keyword_classifying (#5197), registered_cdb_does_not_shadow_a_resolvable_bgsm, unresolvable_bgsm_still_falls_back_to_cdb_pbr, starfield_single_channel_kinds_are_parked_by_name_or_canonical (#4429, anchored to nifal.md's list span by #5003); merge.rs non_colour_slot_replacements_stay_parked (#5190), slot_height_texture_stays_parked_and_never_arms_parallax (#5381), cdb_is_glass_with_alpha_blend_classifies_as_glass_end_to_end (#5277), cdb_emissive_settings_are_parked_not_forwarded_raw (#5283, parked by #5503), alpha_test_requires_an_authored_has_opacity_true (#5196 / #5661), only_a_decodable_loose_mat_is_applied (#5396); index.rs parent_edge_is_decoded_from_the_dbfile_index_and_inherited (#5578: drives the ObjectInfo.Parent decode through the synthetic CDB, with an orphaned control), lookup_inherits_settings_down_the_parent_chain_nearest_wins, lookup_terminates_on_a_parent_cycle (#5502); loose_mat.rs decodes_the_installed_loose_mat_layout, out_of_range_index_is_dropped_not_wrapped, partial_colour_is_not_completed (#5396); reader.rs unit tests (skip_and_read_consume_identically_on_a_reordered_class #5323, hostile counts, pinned vocabularies, read_order_reorders_xmcolor_shape, self_referential_struct_is_rejected_instead_of_overflowing_the_stack); #[ignore]d tests/real_cdb.rs (material_index_resolves_a_vanilla_material, material_index_dlc_cdb_resolves_the_base_corpus, streaming validator — never revert to parse, ~9.19 GB).
Checklist:

• Discovery — discover_starfield_cdbs scans every archive for the base and DLC/Creation materials\creations\<plugin>\materialsbeta.cdb (13 CDBs, ~232 MB; two full-size), peek_magic then the tolerant probe_header (#4273); re-hardcoding the base path drops every DLC CDB.
• Reader — consumption modes: parse (full Value tree; 9.19 GB on ONE full-size CDB, ~18 GB across the set), visit_instances_with_limits (each top-level value delivered then dropped; one LIST/MAPC can still be huge), validate_instances_with_limits (no tree; #4274), and the Phase-2 MaterialIndex::build (crates/sfmaterial/src/index.rs: one streaming pass that fails loudly on degraded joins (missing DBFileIndex, row/instance mismatch, trailing bytes, DIFF/USRD payloads, duplicate index: #5320, 6489ae225, build_rejects_*), re-measured #5291: ~1.6 s, ~247 MB peak, ~115 MB resident per index on the vanilla base CDB (the earlier "~470 MB" is stale); PersistentID → DBID join, material → LayerID → MaterialID → TextureSetID → MRTextureFile slots SLOT_COLOR..SLOT_TRANSMISSIVE = 0..8). The built index is memoised per CDB in sf_cdb_index_cache (cdb.rs) as a single-flight OnceLock slot (CdbIndexSlot; Some(None) memoises a failure, so a failed build can never clobber a concurrent success — the slot replaced 870ea1d07's vacant-only fill), and discover_starfield_cdbs starts each valid CDB's build on a named worker thread so it overlaps the ESM parse (#5291, dd759b762; builds are serialised process-wide so N CDBs never stack N peaks; a lookup arriving mid-build parks on the same slot; guards prefetch_fills_the_lookup_slot_without_a_lookup, prefetch_leaves_an_existing_slot_alone); an unfetchable CDB archive warns before the None is memoized (#5319, 8be8c3d3a, unfetchable_cdb_resolves_none_and_memoizes_the_failure); provider rebuilds reuse opened archive handles (Archive::open_shared, #5288, 5402b2530) and the CDB path scan is memoised on the archive; lookup_cdb_material is last-registered-wins across CDBs. Duplicate class/field names error (#4272); value nesting is capped at 64 (Error::NestingTooDeep, #4657 — an 84-byte self-referential CDB used to overflow the stack); unknown ChunkType/Value variants must error or warn-and-skip, never panic.
• Lookup key (solved, #3398 spike) — reflected CRC-32 (poly 0xEDB88320, init 0, no final XOR) over the lowercased backslash path, directory and stem hashed separately (3,032/3,084 = 98.3%); BSResource::ID labels are rotated (.Dir = stem, .Ext = directory, .File = "mat"). Reproduce: crates/sfmaterial/examples/cdb_key_hash_probe.rs. Text calling the key or field vocabulary "unknown" is stale.
• Field order is offset order (#4275, fixed in 224a19372) — read_user_class walks Class::read_order (offset-sorted), not declaration order; XMCOLOR declares r,g,b,a at offsets 2,1,0,3, so a declaration-order read is a straight R↔B swap on every decoded colour. Regression guard: read_order_reorders_xmcolor_shape.
• Try-then-fall-through (#3230) — the key's extension column is the constant "mat", so lookup ignores the reference's suffix (17 of 57 .bgsm/.bgem-named sample paths resolve to real CDB materials): .bgsm/.bgem names hit resolve_bgsm/resolve_bgem first and reach the CDB only on a miss; .mat keeps its early return, but a loose .mat file actually present in the archives is tried first and, if it decodes, wins over the CDB (apply_loose_mat in merge.rs, #4277 → Merged via apply_cdb_material, recording ImportedTextureSource::Mat). An undecodable one returns None and falls through to lookup_cdb_material and its miss fallback (#5396, faf09c6e6, which also resolved #5436's lost external_material_resolved); the decoder reads Objects[].Components[].Data with string-typed bools/scalars and never completes a partial Color. On every arm a CDB hit goes through apply_cdb_material (Merged) and only a lookup miss falls to apply_cdb_pbr_fallback (PresenceOnly, is_pbr + external_material_resolved). An early PresenceOnly above the resolvers discards every authored role, glass_enabled and PBR scalar.
• Phase-2 state (apply_cdb_material, merge.rs) — slots 0/1/7 fill base_color/normal/emissive with NIF-first fill precedence; slots 2/3/4/5/8 (_opacity/_rough/_metal/_ao/_transmissive, ~39% of Starfield textures) are parked by name in docs/engine/nifal.md (#4429; the guard fails if a role lands without moving the parked row). Slot 6 _height is parked too (#5381, 46dfdcc6c): it is a layer height-blend input (the schema's only POM fields are on ProjectedDecalSettings), and forwarding it into height armed parallaxMapIndex on every surface. EffectSettingsComponent.BlendingMode == "AlphaBlend" → has_alpha (#5277), so IsGlass can reach the glass classifier at all; EmissiveSettingsComponent (Enabled + nested EmittanceSettings) is captured in the CDB lookup but parked — nothing reaches emissive_color/emissive_mult until the physical LuminousEmittance scale is measured (4667b884b, #5503). Settings inheritance through ObjectInfo.Parent is resolved (5c14ebdf8, #5502: IsGlass 4 → 217, emissive Enabled 0 → 619 referenced materials, AlphaBlend still 0). Known open, don't re-file: the remaining half of #5502 (texture-set slots / flat colours are not Parent-inherited: 2,051 of 10,158 referenced materials resolve with no texture), #5509 (additive-family blend modes, 491 shapes, neither forwarded nor parked). MaterialParamFloat stays untranslated (index → roughness/metalness unverified, #3398 open — do not guess); a hit with no translated scalar stamps PbrMaterial::NO_SIGNAL_NEUTRAL so the keyword classifier never fabricates conductors from the resolved colour path (#5197: iris_iron_color.dds → 0.9 metal). TextureReplacement is per-slot (#5190): only the SLOT_COLOR flat colour lands, into diffuse_color, and only when no colour texture bound. IsGlass → bgem_glass (never thin_glass); UseSSS is parked until a subsurface-colour source exists; alpha test requires an authored has_opacity == Some(true) on the Parent chain (#5196; #5661 flipped the absent case — None is the class default false, so 1,265 of 10,152 NIF-referenced materials no longer alpha-test against base-colour alpha; the default is census-measured, not sourced). Any of these reverting is the regression; the parked items are scope, not new findings.
  Output: /tmp/audit/starfield/dim_3.md

Dimension 4: ESM Resolve Rate + Cell Bring-up

Scope split with /audit-esm: it owns the parser as a parser; this dimension owns Starfield's data through it. Shared-mechanism defects → /audit-esm.
Subagent: general-purpose
Paths: byroredux/src/sf_smoke.rs, crates/plugin/examples/{sf_smoke,sf_parse_check}.rs, crates/plugin/src/esm/records/parse.rs, crates/plugin/src/esm/cell/{support,walkers}.rs, byroredux/src/cell_loader/spawn.rs, byroredux/src/systems/light_anim.rs
First step (two different questions): cargo run --release -- --esm Starfield.esm --sf-smoke <CELL_EDID> (per-cell base-form resolve rate, Cydonia) and cargo run --release -p byroredux-plugin --example sf_smoke -- <ESM_PATH> --tsv (GRUP byte coverage vs DISPATCH_HANDLED_FOURCCS, diffed against .claude/audit-baselines/sf-esm/*.tsv).
Guards: DISPATCH_HANDLED_FOURCCS is derived beside the dispatch and pinned in records/tests.rs (#4278); starfield_ligh_dat2_decodes_to_light_data; light_anim.rs starfield_light_type_enum_drives_the_spot_shape + starfield_shadow_technique_follows_the_light_type_enum.
Checklist:

• A resolve-rate drop = the CELL handler dropped REFRs (moved subrecord, new XCLL field) or a base record failed to index → unit-cube placeholders. Check the per-type breakdown for Starfield-only base types (GBFM/GBFT/PNDT/STDT/BIOM); note frequency, don't re-report the known GBFM stub. PDCL (74.9% of unresolved Cydonia REFRs, 2026-08-30) outranks GBFM (0.081%) under the baseline doc's own promote/defer rule. Model-less STAT/BNDS/ACTI/ARMO forms (geometry in a BFCB block) still drop (#1576).
• PKIN — every Starfield PKIN CNAM is a template CELL (11,281 records; 97,981 REFRs place one). #5231's FO4-measured skip-with-warn therefore spawns nothing for any pack-in — open #5489 (Cydonia loses 12,218 template placements, 98.9% STAT). sf_smoke counts those REFRs as resolved through the nominal empty-model StaticObject (open #5533), so the resolve rate cannot see the gap or its fix.
• LIGH — carries no MODL/DATA, only a component-block DAT2; build_static_object_from_subs must decode it or every LIGH REFR misses. Shape is the DAT2+56 Light Type enum (0 omni, 1 shadow spot, 2 non-shadow spot) → LightData::starfield_light_type → translate_light / canonical_light_shadow_flags; flag 0x200 is "Focus Spotlight Beam", not a spot bit (pinned by translate_light_excludes_starfield_even_with_spot_bit_set).
• PDCL stays a named skip (index.skipped_unconsumed_groups + one-shot warn), not the anonymous catch-all. XCLL_SIZES_STARFIELD = [28, 108]: the 108-byte body shares only bytes 0–39 with Skyrim (decoded against xEdit SF1 wbStruct(XCLL); "Skyrim 92 + a 16-byte tail" is stale — two docs still say it, open #5534); parse_lgtm reads Starfield's 108-byte LGTM DATA with the same SF layout (#5002, starfield_lgtm_data_decodes_the_sf_xcll_layout), and the ×70 distance lift of both carriers is pinned through parse_esm by starfield_index_lifts_lgtm_and_xcll_lighting_distances (#5171, 2dab4ff48). Both carriers decode bytes 28..108 through one decode_starfield_lighting_tail (esm/cell/mod.rs) and one starfield_heights lift (spatial_units.rs) (#5346, 9dbd59b8b) — a label correction lands once; a third hand-copied offset table is the regression. XRGD ragdoll-pose offsets are lifted to engine units in spatial_units::cell (#5299, 5f1a862be; no pose consumer yet, #5015). The metric-unit lifts of Starfield WTHR FNAM (#5001/#5134; only AUTHORED distances lift — WeatherRecord::fog_distances_authored, starfield_wthr_without_fnam_keeps_the_unlifted_fog_defaults, #5301, ec75b1439) and WATR DNAM (#5151, pins corrected in 24cb577d2; three unit-unsettled lane families deliberately unlifted, pinned by starfield_watr_defers_the_unclassified_dnam_lanes, #5170) happen at the parse boundary and belong to /audit-exterior, as does WTHS, Starfield's WTHR replacement (#5364, a5484faf6: the BGS reflection blob's schema half is decoded, its instance values are censused but not interpreted). TXST decode covers TX00–TX07; unmodelled TX08/09/17/19 (metal/rough/ao/opacity) warn once per FourCC (warn_unmodelled_txst_slot, #4438) — capture waits on canonical roles (Dim 3).
• Spawn gates (cell_loader/spawn.rs) — static-trimesh fallback gated on base_layer, not final_layer (#1294); synthesized colliders (spawn_trimesh_collider_ghost / spawn_packed_collision_proxy) carry no MeshHandle, so they never enter blas_specs; DoorTeleport from REFR XTEL (#1295).
  Output: /tmp/audit/starfield/dim_4.md

Dimension 5: NIF Shader Blocks — BSVER 155+

Subagent: legacy-specialist
Paths: crates/nif/src/blocks/shader/{mod,lighting,effect}.rs, crates/nif/src/blocks/shader_tests/starfield.rs, crates/nif/src/shader_flags.rs
First step: cargo test -p byroredux-nif --lib -- shader::tests::starfield — the test module mounts via #[path = "../shader_tests/mod.rs"] mod tests inside blocks::shader, so a shader_tests::starfield filter matches zero tests (vacuous green, #5004)
Guards: starfield.rs — parse_bs_lighting_starfield_captures_trailing_tail, the ..._tail_empty_without_size_or_drift pair (LSP + effect), every_tail_capturing_block_reports_it_and_parse_nif_records_it (#2532). NIF mechanics are /audit-nif.
Checklist:

• CRC32 flag arrays for BSVER ≥ FO4_CRC_FLAGS (132) → sf1_crcs; SF2 for BSVER ≥ FO76_SF2_CRCS (152) → sf2_crcs; hashes are the same reflected CRC-32 as CSG/CDB, over the uppercase nif.xml flag name (bs_shader_crc32). The effect-shader Own_Emit additive promotion reads typed bit or the CRC EMIT_ENABLED flag via modern_effect_shader_bit (#4279, own_emit_additive_promotion_reads_the_typed_bit_and_the_crc_flag); no vanilla Starfield block sets it, so a typed-word-only read is invisible on vanilla.
• #1510 — the BSShaderType155 tail once over-read by 4 B, truncating ~1,036 full-body BSLightingShaderProperty blocks to NiUnknown; that count must stay 0.
• Undocumented tails — empty-name full-body BSLightingShaderProperty and BSEffectShaderProperty carry trailing bytes nif.xml does not document; both capture block_size - consumed opaquely into starfield_tail via read_starfield_tail (never a hardcoded length; LODMeshes drift 0). Never fabricate field names or semantics.
• Material-reference stubs — a non-empty Name at bsver >= STARFIELD is a stub (is_material_reference): census of 480,861 = 478,691 .mat, 1,679 .bgsm, 104 .bgem, **387 suffix-less, all the degenerate Materials\ / \Materials** (editor markers, conveyors — not content-hash paths, so the "hash path" comments/fixtures are wrong, #4439). Those get material_path = None → Unresolved, the only stubs that never reach the CDB PBR route.
  **Output**: /tmp/audit/starfield/dim_5.md

Dimension 6: Material Flow — NIFAL Boundary + BGSM/BGEM/.mat

Subagent: renderer-specialist
Paths: byroredux/src/material_translate.rs, byroredux/src/asset_provider/material/merge.rs, byroredux/src/cell_loader.rs (pack_imported_material_flags), crates/nif/src/import/material/slot_role.rs, crates/bgsm/src/{bgem,bgsm}.rs
First step: git log --since=<last report> --format='%h %cs %s' -- byroredux/src/material_translate.rs byroredux/src/asset_provider/material crates/nif/src/import/material
Guards: merge_external_material_is_the_only_exported_fn_in_this_file; colocated_lighting_mask_is_confined_to_the_tint_family_slot_two (#4431); bgsm_merge.rs / starfield_mat.rs. Canonical boundary invariants are /audit-nifal.
Checklist:

• merge_external_material takes &mut ImportedMaterial (cannot touch geometry/skinning); .mat texture paths must land in MaterialTextureSet roles, never a CDB slot index; BGEM stays distinct from BGSM (glass_enabled, the authoritative glass signal, must not misclassify an opaque piece with a stuck flag). pack_imported_material_flags derives BGSM_AUTHORED / PBR_BSDF / TRANSLUCENCY / MODEL_SPACE_NORMALS / EFFECT_PALETTE_COLOR from the right fields; on Starfield BGSM_AUTHORED, TRANSLUCENCY and MODEL_SPACE_NORMALS can never be set today (zero BGSM files). from_bgsm stays FO4 spec-glossiness (both CDB arms leave it clear); effect-shader glass promotion reads the separate external_material_resolved (set by the BGSM/BGEM arms and both apply_cdb_material / apply_cdb_pbr_fallback, #4283) in classify_glass_into_material_with_provenance, while the 2..=20 authored-dispatch guard deliberately still keys on from_bgsm (the CDB authors scalars, not the dispatch). Swapping either signal is the regression.
• Resolve-once — BGEM and Starfield stubs whose CDB lookup misses leave metalness/roughness NaN so resolve_pbr's keyword classifier runs (has_no_pbr_classifier_signal) — a live arm, not a "future backstop" (doc corrected in c2b67d81e, #4441); a CDB hit stamps the no-signal neutral instead (#5197, Dim 3). Starfield uses FO76's slot vocabulary (#3900); slot_to_colocated_role is Skyrim-only (#4431). Sinks (#4282): wetness/luminance reach Material as WetnessShading/LuminanceShading — capture-only, no GpuMaterial consumer yet. Water concentration normalises once at the WATAL boundary (resolve_water_material, #4285); a per-game token back in water.frag is the regression.
• Population fact (2026-08-30 block histogram) — vanilla Starfield ships zero NiPSysEmitter*, BhkMultiSphereShape and BhkConvexListShape blocks, so the NIFAL particle and per-shape collision slices have no Starfield input and a test there is vacuous; a non-zero count in a new archive is a finding. Collision is BhkSystemBinary blobs; Cydonia's colliders are synthesized (Dim 4).
  Output: /tmp/audit/starfield/dim_6.md

Phase 3: Merge

1. Read /tmp/audit/starfield/dim_*.md; combine into docs/audits/AUDIT_STARFIELD_<TODAY>.md:
   • Executive Summary — first-class GameKind; NIF + BA2 at the measured rate; CDB Phase 2 partial (roles + flags translated, scalars, slot-6 height and five kinds parked); walkable Cydonia; regressions in the bring-up surface.
   • Dimension Findings by severity; CRC32 Flag Table (flag → CRC32 via bs_shader_crc32, read-by-import, vanilla occurrences).
   • Remaining-Work Chain (starfield-esm-roadmap.md: Phases 0+1 done, 2–4 invalidated by the 99.9%-parity measurement) — CDB Phase 2 remainder (#3398: the indexed reader, XMCOLOR offset order, IsGlass signal, AlphaBlend + emissive-settings capture (#5277/#5283) and the real-layout loose .mat decoder (#5396) shipped; left are canonical roles for the five parked kinds (#4429), a height-blend consumer for slot 6 (#5381), MaterialParamFloat → roughness/metalness mapping, a subsurface-colour source for UseSSS, texture-set slot Parent inheritance (rest of #5502), the additive-family blend modes (#5509)) → PKIN template-CELL instancing (#5489) and PDCL ahead of GBFM → exterior worldspace tiles → space-cell / planet / GBFM records. Never frame it as "BGSM parser first / ESM very far" — both shipped; the NIF truncation tail is cleared.
2. Remove cross-dimension duplicates.

Suggest: /audit-publish docs/audits/AUDIT_STARFIELD_<TODAY>.md
(label every finding game:starfield + legacy-compat, plus its own domain label.)