feat(corpus): third witness IFClite and the full-project export artifact g-2026-0002

[DrunkOnJava]

Syncs the umbrella with rvt-rs #205 (dbf2347) and #207 (bdddd52).

The edge is still one edge — RVT → IFC via the Revit 2024 exporter — but it now carries two artifacts and three independent reading lineages, and the second artifact is the one that says how much the first was worth.

The third witness: IFClite

ifc-lite is adopted: crate ifc-lite-core 7.1.1 (MPL-2.0, Rust, upstream LTplus-AG/ifc-lite), a byte-level STEP scanner sharing no code with IfcOpenShell. That resolves the registry's one open entry, which carried a wrong license (MIT) against a bare ifc-lite slug and was gate-ineligible under SPEC.md §19 note 4. registry/README.md now records what it resolved to instead of what was unknown.

witnesses/ifc-lite/ vendors the glue from rvt-rs tools/ci/witness-ifc-lite/ at dbf2347. Only the manifest-reading path differs: it takes the bridge file directly and checks bridge.file_hash_sha256 / bytes / schema, the same contract tools/witness-ifcopenshell.py already implements. Counting, canonicalization, hashing and the observation shape are unchanged — which is why the committed observations still hash identically across both repositories. It is its own Cargo workspace root, so MPL-2.0 code never links into the Apache-2.0 tree. It is glue, not a decoder: the umbrella still parses no format bytes itself.

g-2026-0002: the full project export

Same .rvt, same committed rvt-rs observation, different bridge file: 2024_Core_Interior_slim.ifc, 1,665,968 bytes, 19,879 entities, against g-2026-0001's 20 KB element fixture. The origin URL was verified by fetching and hashing.

The result is the honest one:

Category	Export	rvt-rs	Status
walls (IFCWALL)	360	0	known_gap, rvt-rs#30
doors (IFCDOOR)	132	0	known_gap, rvt-rs#32
columns (IFCCOLUMN)	256	0	known_gap, rvt-rs#204
rooms/spaces (IFCSPACE)	116	18	known_gap, rvt-rs#33
floors (IFCSLAB)	80	64	known_gap, rvt-rs#31
levels (IFCBUILDINGSTOREY)	15	12	decoder_baseline, rvt-rs#33
windows (IFCWINDOW)	6	0	known_gap, rvt-rs#32
materials (IFCMATERIAL)	10	102	known_gap, rvt-rs#34 — over-count
property sets (IFCPROPERTYSET)	0	64	known_gap, rvt-rs#35 — over-count against zero

So the claimed surface is four fields and nine categories are excluded first-class with their tracking issues. The excluded table is the deliverable, not a footnote. levels is retained here though it is dropped in g-2026-0001: there it has no source_ifc_type, here it has one and both bridge readers see it.

Gate

verdict.yml installs Rust stable (dtolnay/rust-toolchain, SHA-pinned), caches with Swatinem/rust-cache, builds the glue with --locked and runs it for every artifact alongside IfcOpenShell. A manifest bridge witness now declares runner_kind (python or cargo); tools/replay.sh dispatches on the same field.

tools/check-registry.py gained a §9.6 pin check for anything the umbrella builds itself: the registry version, the exact Cargo pin, a committed Cargo.lock, the manifest version_pin and the binary's own WITNESS_ID/WITNESS_VERSION must be one number. That mirrors rvt-rs's ifc_lite_gate_is_wired_and_version_pinned, which cannot be imported here.

SPEC 1.0.1

SPEC.md is a byte-identical copy of rvt-rs docs/octetproof-spec.md at bdddd52. Patch release, non-semantic: §19 note 4 records the ifc-lite resolution, and the header and §16.3 point at this repository instead of calling it planned. No schema, diff-function or provenance change, so every document here still declares schema_version: "1.0.0".

Local verification

$ python3 tools/validate-corpus.py
corpus: 6 observation(s) and 2 verdict(s) conform to the 1.0.0 schemas

$ python3 tools/check-registry.py
witnesses.json: valid — 11 nodes, 47 witnesses, 7 artifacts, 3 edges, 9 agreements; 2 corpus manifest(s) resolve

$ python3 tools/index.py --check
corpus/MANIFEST_INDEX.json: chain head 7a8a62ca4449… ok (2 artifact(s))

$ tools/replay.sh corpus/artifacts/g-2026-0001
g-2026-0001: PASS — witnesses ifc-lite, ifcopenshell, rvt-rs
  surface: 8 fields, excluded: 4, diffs: 0
  replay ifc-lite: match
  replay ifcopenshell: match
  replay rvt-rs: match
verdict matches the committed record (ignored: artifact_id, replay, timestamp, verdict_hash_sha256)

$ tools/replay.sh corpus/artifacts/g-2026-0002
g-2026-0002: PASS — witnesses ifc-lite, ifcopenshell, rvt-rs
  surface: 4 fields, excluded: 9, diffs: 0
  replay ifc-lite: match
  replay ifcopenshell: match
  replay rvt-rs: match
verdict matches the committed record (ignored: artifact_id, replay, timestamp, verdict_hash_sha256)

Both bridge observations hash identically to each other on both artifacts (8cf40465… and 882e1e0f…): two unrelated STEP readers, different languages, byte-identical canonical payloads. That corroborates the bridge-side reading and nothing more — entity_counts is a shallow surface.

Chain head 7a8a62ca4449a3e22c31dfbb6eff44fe73a0fde3f776d9f1ab7c279e99279842; g-2026-0002's prev_hash is the canonical hash of g-2026-0001's entry.

https://claude.ai/code/session_01BaHKf1Rw5aJBGK5y3xmx7C

---

Note

Medium Risk
Changes expand the fail-closed CI gate (Rust builds, crates.io/PyPI pins, corpus verdict hashes) and license isolation for MPL glue; mistakes in pinning or runner dispatch would break replay for all artifacts.

Overview
Adds a third gated witness (IFClite ifc-lite-core 7.1.1, MPL-2.0) and a second golden artifact (g-2026-0002) on the same RVT→IFC edge, so the corpus now exercises three lineages and contrasts the thin 20 KB IFC fixture with the full ~1.6 MB project export.

IFClite is adopted in registry/witnesses.json (upstream/license resolved to LTplus-AG/ifc-lite / MPL-2.0). New witnesses/ifc-lite/ is separate-workspace Rust glue (pinned Cargo.lock, --locked builds) that emits §6.2 observations; g-2026-0001 gains runner_kind and an ifc-lite bridge entry plus committed observation and recomputed verdict. g-2026-0002 adds manifest, observations, verdict, PROVENANCE.md, and an updated MANIFEST_INDEX chain.

Gate and replay now dispatch bridge witnesses by runner_kind (python vs cargo): verdict.yml installs Rust stable, caches witnesses/ifc-lite, builds cargo runners, and runs both bridge processes; tools/replay.sh mirrors that. tools/check-registry.py enforces §9.6 pin alignment for cargo witnesses (registry version, Cargo.toml, manifest version_pin, WITNESS_* in main.rs).

Docs/spec: SPEC.md 1.0.1 (patch: umbrella link, ifc-lite license note); README/CONTRIBUTING/NOTICE/corpus docs updated for two artifacts, three witnesses, and replay requiring Rust ≥ 1.87.

^{Reviewed by Cursor Bugbot for commit 278a2c4. Configure here.}

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_fdfb4432-3b47-47e5-aff0-620e0e31e8f8)