LOG

Running log of noteworthy work; newest first.

2026-07-17 21:23 — Complete the port: hardened, reviewed, all gates green

What: Hardening + adversarial-review workflow finished: fixed both fuzz crash classes (unchecked Int(UInt64) on hostile size fields in BinaryParser; negative KeyCount → Swift Range trap in TakesReader, fixed at ufbx's exact "Z"-format rejection boundary, ufbx.c:7743) and applied 8 verified review findings — notably f64ToI64 trapping at the 2^63 boundary (Double(Int64.max) rounds UP to 2^63, admitting a value that traps; use exclusive < 9223372036854775808.0) and ufbx's 0.99999f layer-weight clamp needing Double(Float(0.99999)) for bit parity.
Why: A parser must throw, never trap, on malformed input; the review lenses caught boundary bugs fuzzing and goldens couldn't reach.
Impact: Final state: 144 tests / 0 failures (44/44 golden, 44 malformed-truncation cases), zero crashes over 749 corpus files + 2400+ mutations on debug AND release, clean builds. The "release exits 0 / debug traps" mystery was a fuzz-harness artifact (shared temp path saved post-hoc), not a code bug — verify crasher inputs are saved at crash time before chasing such ghosts. Tree is complete but uncommitted.

What: Closed the last two golden classes: (1) pre-7000 frame rate lives in Version5 > Settings > FrameRate, which ufbx turns into synthetic CustomFrameRate+TimeMode=CUSTOM props (ufbx.c:15766/15921); (2) 6100 textures wire via LayerElement*Textures mesh blocks + a patch-from-meshes pass (ufbx.c:22364–22466). Then fuzzed: 749 corpus files + truncations + seeded mutations → 7 crashers (unchecked int conversion in BinaryParser.parseNode + a nondeterministic release-mode class). Hardening/review workflow launched.
Why: Golden parity is the port's acceptance bar; crash-on-malformed-input is unacceptable for a parser library.
Impact: Gotcha found en route: ufbx's dst-sorted connection array tie-breaks on ORIGINAL file order, not src-sorted order (ufbx.c:18768–18774) — was silently wrong for 42/44 files because ties are rare; TextureId indexing exposed it. Fuzz repro harnesses live in the session scratchpad (fuzzrun2.py, seed 42); crash inputs preserved to Tests/Resources/malformed/ by the running workflow.

What: All six implementation waves completed (19 agents, zero failures): full library from inflate/float-parse through parsers, scene model, readers, linker/finalizer, evaluation, triangulation, and the golden dumper. Package builds clean on Swift 6; every unit suite passes; first-ever golden run scored 31/44.
Why: Milestone — the port is functionally complete end-to-end; remaining work is parity debugging, not construction.
Impact: The 13 golden failures cluster into 3 classes (spurious face_material arrays ×11, 6100-ASCII texture→material links, Max time_mode/fps mapping) — a parity fix loop is running. Gotcha for future integrations: cross-wave interface drift was ~zero; the only integration fixes needed were two test-file bugs, validating the contracts-first DESIGN.md approach.

What: Two adversarial reviewers audited docs/DESIGN.md against the ufbx notes before any code was written; 18 findings fixed. Then launched the 6-wave implementation workflow (foundations → parsers+model → readers → link/finalize → eval/dump → integrate) with strict per-agent file ownership.
Why: Worst finding would have silently broken every 6100 file: synthetic fbx_ids apply to ALL version < 7000 objects (not just <6000) — without them every connection dangles and the scene graph comes out empty. Also: Takes must be read for 7x00 too (stack time-range back-fill), and InheritType maps non-ordinally (0→componentwiseScale, 2→ignoreParentScale, else normal).
Impact: These traps are now recorded in DESIGN.md's "Known fidelity traps"; check that list first when a golden diff shows empty scenes, degenerate anim ranges, or wrong scale inheritance.

What: Scaffolded the SwiftPM package and built the verification backbone before any parser code: vendored ufbx (rev in tools/ufbx/VERSION), wrote a C harness (tools/ufbx_dump.c) emitting a deterministic JSON scene dump (spec: docs/DUMP_FORMAT.md), generated golden dumps for 44 representative FBX files, and fanned out 15 agents to write line-cited port notes (docs/ufbx-notes/).
Why: Porting 33k lines of quirk-laden C needs an executable acceptance criterion, not code review by vibes — structural parity with real ufbx output (1e-6 float tolerance) catches subtle behavioral drift per file/feature.
Impact: swift test golden suite defines "done" for the port. Key design decision recorded in docs/DESIGN.md: ARC-safe arena ownership (scene owns elements; cross-refs are Int32 indices behind computed properties), mirroring ufbx's typed_id model to avoid retain cycles.