Asset-reuse parity: canonical glb names, shared assets layout, pre-build probe

[dalkia]

What

Brings the scene path (live/JIT server + abgen-corpus) to parity with the Unity converter's asset-reuse flow, which the ab-cdn deployment has run since v49. Previously abgen tracked each glb's dependencies only inside the bundle payload (metadata.json / m_Dependencies) — dependency hashes never reached bundle names, S3 keys, or manifests.

On by default (matching the v49+ deployment); ABGEN_ASSET_REUSE=0 opts out for parity runs against pre-v49 reference trees.

Changes

• Canonical naming — scene glb/gltf bundles are named {hash}_{depsdigest}_{platform} (textures stay {hash}_{platform}), so a glb whose dependency set changes lands at a new name/key. The JIT name parser accepts three-part names and validates the requested digest against the one computed at entity scan (mirrors the C# throw on missing digest).
• Digest parity fix — compute_deps_digest now filters to GLB_DEP_EXTENSIONS (.bin + textures), mirroring upstream computeDepsDigest. Previously abgen hashed all resolved refs, diverging for glbs referencing e.g. .ktx2. Also corrects the wearables path.
• Shared S3 layout — uploads move from entity-scoped {version}/{cid}/{file} to upstream's {version}/assets/{canonical_name}; reads try the shared key first with entity-scoped fallback for pre-switch objects.
• Pre-build probe — new signed Space::head; corpus builds HEAD-probe the canonical key and list hits in the entity manifest without rebuilding (upstream checkAssetCache behavior).
• Serving — the abcdn flat lane strips the digest before owner-entity resolution so canonical-name requests JIT correctly.
• Missing-dep semantics — strict mode skips the glb (upstream skipped-assets behavior); ABGEN_MAGENTA_MISSING drops unresolvable deps from the digest and builds with placeholders.

Verification

• Full workspace green (329 lib tests + suites), clippy clean.
• New tests: digest extension filter, canonical-stem parsing, shared-layout put/get ordering with legacy fallback, HEAD probe hit/miss/legacy-off, corpus derive in legacy/reuse/strict/tolerant modes.
• End-to-end with the real binary: default run emits Qmglbtest_4f53cda18c2baa0c0354bb5f9a3ecbe5_windows (byte-identical to the TS empty-deps digest vector) with canonical names in the manifest; ABGEN_ASSET_REUSE=0 reproduces today's Qmglbtest_windows output exactly.

Notes for review

• --live-mode sampling of pre-v49 vintages needs ABGEN_ASSET_REUSE=0 to intersect with those reference manifests (documented in README). Per-vintage naming inside from_live_reference is a possible follow-up.
• Probes run serially per entity (one HEAD per asset on first JIT build); parallelizing is a small follow-up if it shows in build latency.

🤖 Generated with Claude Code