corpus: mirror rvt-rs #226/#230/#231 — relations and storeys join the surface; spec 1.1.0

[DrunkOnJava]

Brings the umbrella to OctetProof 1.1.0 and mirrors the current rvt-rs corpus state (rvt-rs main 5da2570, PRs #226 / #230 / #231). The protocol gains two additive field classes; the corpus gains two agreement classes it did not have — topology, and labels-with-units.

The result

Artifact	Before	After
g-2026-0001 (20 KB element fixture)	PASS, 4 surface fields, 9 excluded	PASS, 6 surface fields, 10 excluded
g-2026-0002 (full project export)	PASS, 10 surface fields, 4 excluded	PASS, 13 surface fields, 3 excluded

g-2026-0002 now claims, agreed exactly by all three witnesses at tolerance 0: 360 IFCWALL, 256 IFCCOLUMN, 132 IFCDOOR, 80 IFCSLAB, 20 IFCSHADINGDEVICE, 15 IFCBUILDINGSTOREY, 6 IFCWINDOW, 1 IFCUNITASSIGNMENT, three zeroes — plus 138 IfcRelFillsElement [host Tag, filling Tag] pairs and the 15 [name, elevation] storeys, Basement 2 at −40 ft through Level 13 at 185.5 ft.

Three known_gap exclusions remain, and they are the deliverable: IFCSPACE 116 vs 18, IFCMATERIAL 10 vs 102 (an over-count), IFCPROPERTYSET 0 vs 854 (an over-count against zero). That table used to have nine rows.

g-2026-0001 stays thin — three of its five counts are zero on both sides — with one exception worth recording: Revit writes the complete fifteen-storey spatial hierarchy into even a one-element 20 KB export, so storeys.IFCBUILDINGSTOREY is as strong on the fixture as on the 19,879-entity project export. relations.IFCRELFILLSELEMENT goes the other way and is excluded decoder_baseline: that fixture carries no IfcRelFillsElement at all, so the two sides of that edge are not comparable on it.

What changed

Protocol — SPEC.md 1.0.2 → 1.1.0 and both schemas, byte-identical copies from rvt-rs (the same policy as the registry). schema_version now an enum of ["1.0.0","1.1.0"]; optional observation.relations / observation.storeys; verdict diffs gain only_in_a / only_in_b. Minor and additive (§16.1, §20) — the diff function, canonicalizer, provenance model and status vocabulary are unchanged, and a 1.0.0 observation is still valid input. registry/witnesses.json was already byte-identical and is untouched.

Gate — tools/verdict.py ports both field classes from tools/ci/witness-verdict.py: exact sorted-multiset equality, no tolerance concept, diffs carrying set sizes plus the members each side holds alone. The surface is built counts → relations → storeys, the order the committed verdict records. A witness compared on a set-valued class must declare it in semantic_surface_covered or the verdict is MANIFEST_ERROR. The umbrella-manifest shape is kept — semantic_surface / excluded stay explicit and now cross-check against all three blocks. tools/witness-ifcopenshell.py ports the extraction, resolving the elevation with ifcopenshell.util.unit. tools/jsonschema_mini.py implements maxItems, the one keyword the new schemas introduce — it fails loudly on unimplemented keywords, so without it the corpus would not validate at all. compare-verdict.py, fetch.py, index.py, check-registry.py and replay.sh needed nothing.

Witness glue — witnesses/ifc-lite/ re-mirrored from rvt-rs a0044f4 (PR #230), gaining fills_element_pairs and building_storey_set (units via ifc_lite_core::extract_length_unit_scale). The only deviation from upstream remains the manifest-reading path. Cargo.lock unchanged, no new dependency, §9.6 pin still one number: ifc-lite-core = "=7.1.1".

Corpus — observations and verdicts copied verbatim from research/witness/<alias>/; both manifests rebuilt from the project-count fixtures under the documented drop rules, gaining relations and storeys blocks and levels as known; both declare octetproof_version: "1.1.0" because both now carry 1.1.0 blocks. MANIFEST_INDEX.json regenerated, chain head 5e6a64228083.

Local verification — clean _replay, both artifacts end to end

g-2026-0001: PASS — witnesses ifc-lite, ifcopenshell, rvt-rs
  surface: 6 fields, excluded: 10, 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)

g-2026-0002: PASS — witnesses ifc-lite, ifcopenshell, rvt-rs
  surface: 13 fields, excluded: 3, 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)

plus tools/index.py --check (chain head ok), tools/check-registry.py (47 witnesses, 2 manifests resolve, §9.6 pin held) and tools/validate-corpus.py (6 observations + 2 verdicts conform to the 1.1.0 schemas).

Both bridge witnesses hash identically on both artifacts — 99e6cd7a… on g-2026-0001, 5a101408… on g-2026-0002 — across all 14 entity types, all 138 relation pairs and all 15 storey pairs, including the unit resolution, which each reader performed with its own code.

Fail-closed paths exercised, not assumed. Mutating one relation host tag and one storey elevation in a fresh ifc-lite observation produces DISAGREE with only_in_a / only_in_b naming exactly the changed members; stripping storeys from a witness's semantic_surface_covered produces MANIFEST_ERROR; a three-element pair and an out-of-enum schema_version are both rejected by jsonschema_mini.

Docs

README's "what exists today" rewritten around the three agreement classes with the real numbers; both PROVENANCE.md files rewritten (witness-set hashes, surfaces, exclusion tables); corpus/README departure table updated for the 1.1.0 blocks and the octetproof_version policy; tools/README gains the field-class table and the reasons a storey elevation travels as a string and storey names are deliberately not Unicode-normalised; CONTRIBUTING's coverage vocabulary extended.

---

Note

Medium Risk
Touches the verification gate’s diff logic and committed corpus verdicts; mistakes could pass wrong agreement or break CI replay, but the change is additive and schema-backward-compatible.

Overview
OctetProof 1.1.0 adds two comparable surface classes — relations (sorted host/filling pair multisets) and storeys (sorted [name, elevation-in-feet] sets with per-file unit normalization) — and this PR lands the spec, schemas, gate, witnesses, and golden corpus on that release.

Protocol & CI: SPEC.md and observation/verdict JSON Schemas move to 1.1.0 (schema_version accepts 1.0.0 and 1.1.0; optional observation.relations / observation.storeys; verdict diffs may include only_in_a / only_in_b). .github/workflows/verdict.yml and docs now validate against 1.1.0 schemas.

Gate & witnesses: tools/verdict.py diffs the new set-valued fields with exact multiset equality and requires witnesses to declare relations / storeys in semantic_surface_covered or fail MANIFEST_ERROR. tools/witness-ifcopenshell.py emits the new payload keys; tools/jsonschema_mini.py gains maxItems for pair-shaped arrays. witnesses/ifc-lite/ is re-mirrored from rvt-rs (PR #230) with relation and storey extraction.

Corpus: Both artifacts declare octetproof_version: "1.1.0", gain relations / storeys manifest blocks, and ship regenerated observations and verdicts (updated hashes; MANIFEST_INDEX chain head refreshed). g-2026-0002 widens the claimed surface (e.g. walls, doors, slabs, 138 IFCRELFILLSELEMENT pairs, 15 storeys; exclusions shrink from nine gaps to three). g-2026-0001 adds storey agreement on the thin fixture while excluding the relation field where the bridge has no fills.

^{Reviewed by Cursor Bugbot for commit 9ef497e. 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_02b93722-8fdd-45e7-98d0-2e57a6d796d0)