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

.github/workflows/verdict.yml

@@ -71,8 +71,8 @@ jobs:
       - name: Validate the witness registry against its schema
         run: python3 tools/check-registry.py
       # §6.2/§6.3: every committed observation and verdict conforms to the
-      # 1.0.0 schemas, and every committed hash is self-consistent.
-      - name: Validate the committed corpus against the 1.0.0 schemas
+      # 1.1.0 schemas, and every committed hash is self-consistent.
+      - name: Validate the committed corpus against the 1.1.0 schemas
         run: python3 tools/validate-corpus.py
 
   verdict:
@@ -204,7 +204,7 @@ jobs:
 
       # §6.2: the freshly produced observations must conform to the schema,
       # not merely hash to the committed value.
-      - name: Validate fresh observations against the 1.0.0 schema
+      - name: Validate fresh observations against the 1.1.0 schema
         run: |
           python3 tools/validate-corpus.py \
             --extra-observations "_work/${{ matrix.artifact }}/observations"

CONTRIBUTING.md

@@ -33,10 +33,14 @@ Your entry must declare:
 And, in the pull request body:
 
 - **Coverage declaration.** Which fields of the controlled vocabulary you
-  claim. Today that vocabulary is `entity_counts`; `layer_topology`,
+  claim. Today that vocabulary is `entity_counts`, `relations` and `storeys`
+  (the last two added in 1.1.0, SPEC.md §9.4, §20); `layer_topology`,
   `linework`, `bounding_boxes`, `xdata_fields` and `text_content` are
   specified but not yet exercised. Declaring a field you cannot actually
-  parse is a registration violation and grounds for removal.
+  parse is a registration violation and grounds for removal — and the gate
+  enforces the declaration for the set-valued classes: a witness that does
+  not declare `relations` or `storeys` makes the verdict `MANIFEST_ERROR`
+  rather than being quietly skipped.
 - **A version pin.** An exact git SHA or release tag, or the exact PyPI /
   crates.io / NuGet specifier the CI gate will install. Floating ranges are
   forbidden. A witness update is a separate pull request that includes the

corpus/MANIFEST_INDEX.json

@@ -5,26 +5,26 @@
       "artifact_id": "g-2026-0001",
       "bridge_sha256": "d07c7462aee22640661faed5262cf802ce0fcbc663f312961a39be92bf857050",
       "manifest_path": "corpus/artifacts/g-2026-0001/manifest.json",
-      "manifest_sha256": "6620d925a7a89688b991d4b1a8e10ead6191e73b4242502b8b014cdfbc0c3971",
+      "manifest_sha256": "ab33a0f4918587504d083551f3eefef8ec33065268355f9d5f9b487a6436d203",
       "prev_hash": null,
       "source_sha256": "c805df445d613b408e37337765572021265e3f5dfdc7d1fa53b22ba1600b8014",
-      "verdict_hash_sha256": "94d723919686963121c05c7475c3ea126e9f8d97fa4b9136c6e5656421c8b40a",
+      "verdict_hash_sha256": "c615cc6c4606c3486be2a82af7566b4e6a5002623ffd39227f591b9df5182efa",
       "verdict_status": "PASS"
     },
     {
       "alias": "magnetar-2024-core-interior-slim",
       "artifact_id": "g-2026-0002",
       "bridge_sha256": "bfdf36ffb0bb768f3409d818403990e64d4c262c6780603be87f8077387ad86d",
       "manifest_path": "corpus/artifacts/g-2026-0002/manifest.json",
-      "manifest_sha256": "f22783cd973aa9fb4968d72c4b0b47849f4da97574e0c71dcca36de37c39893a",
-      "prev_hash": "6e2a42634621ed26b655dbf1c3bc7cd1488650779e65cf02223a33c66450279f",
+      "manifest_sha256": "76e218f0f74ea1d2172727240d28422238856dc38b70a20d1f62a7624915b2f8",
+      "prev_hash": "65ee5d9750b2225244ea4e2ae873fc5f6a95ae97b000c9ed82feef1412a9499e",
       "source_sha256": "c805df445d613b408e37337765572021265e3f5dfdc7d1fa53b22ba1600b8014",
-      "verdict_hash_sha256": "48ef575a11505ee1fcbb02b37b3551e0e8d5618a1ec39aeb6d6cb250e085648c",
+      "verdict_hash_sha256": "cd171b3167c602d0882dd8a0d38c99791012a9fe693043e43e2d43b028554325",
       "verdict_status": "PASS"
     }
   ],
   "chain_algorithm": "sha256 over JCS-lite canonical entries (sorted keys, no whitespace, UTF-8)",
-  "chain_head": "50fb3adb759b3d2ddb45a9b4e7d33a4c486cba749901777e7f58df1ec3aa7a58",
+  "chain_head": "5e6a64228083a30b9c3ef667d0d6e02ad96b370954b34599c1da18850fcdad23",
   "schema_version": "1.0.0",
   "signature": null,
   "signing_status": "not implemented — SPEC.md §12.2 requires an Ed25519 signature over the chain head; no maintainer key is published yet, so this chain is tamper-evident but not cryptographically anchored"

corpus/README.md

@@ -9,7 +9,8 @@ corpus/
   MANIFEST_INDEX.json          hash chain over every manifest
   artifacts/
     g-2026-0001/
-      manifest.json            SPEC.md §6.1 — hashes, origins, counts, surface
+      manifest.json            SPEC.md §6.1 — hashes, origins, counts,
+                               relations, storeys, surface
       PROVENANCE.md            what is recorded and what is not
       observations/            SPEC.md §6.2 — one per witness, committed
       verdict.json             SPEC.md §6.3 — the recorded decision
@@ -68,35 +69,41 @@ driven by the umbrella's different job:
 |---|---|
 | `artifact_id` + `alias` instead of `id` | the umbrella assigns `g-YYYY-NNNN`; `alias` keeps the id the decoder repository uses, so observations from either side resolve. See the artifact's PROVENANCE.md. |
 | `source` / `bridge` blocks carry `origin`, `license` and `bytes` | the corpus fetches by hash instead of committing bytes; without an origin an artifact is not replayable |
-| the `decoder_*` fields are dropped | they are the decoder repository's own regression baseline, which the gate explicitly ignores (§6.1). A category that exists *only* as a decoder baseline goes with them — `levels` in g-2026-0001's upstream, which carries no `source_ifc_type` and so no bridge witness can see it. Where the same category does carry a `source_ifc_type`, as `levels` does in g-2026-0002, it stays: it is a real cross-witness observation (15 exported storeys against 12 recovered) and is excluded first-class rather than dropped. Dropping it would hide a measured disagreement behind a formatting rule |
-| `octetproof_version` names the protocol document each manifest was authored against | g-2026-0001 says `1.0.0`, g-2026-0002 says `1.0.1`. 1.0.1 is a patch release with no schema, diff-function or provenance change (SPEC.md §19a), so the two conform identically; the field records history rather than a difference in obligations, and an immutable manifest is not rewritten to chase a patch number |
-
-Everything the gate reads is unchanged: the normative `counts` block, with
-`source_ifc_type`, `status` (`known` / `known_gap` / `unsupported`),
-`tolerance`, `tracking_issue` and `unsupported_feature` carrying exactly the
-§6.1 meanings.
+| the `decoder_*` fields are dropped | they are the decoder repository's own regression baseline, which the gate explicitly ignores (§6.1). A category that exists *only* as a decoder baseline goes with them — one that carries no `source_ifc_type` (or, in the 1.1.0 blocks, no `relation_ifc_type` / `storey_ifc_type`) is invisible to every bridge witness and is not a cross-witness category at all. A category that *does* carry its type key stays even when its status excludes it: `relations.IFCRELFILLSELEMENT` on g-2026-0001 is `decoder_baseline` and is kept, because dropping it would hide behind a formatting rule the fact that the two sides of that edge cannot be compared on it |
+| `octetproof_version` names the protocol document each manifest was authored against | both say `1.1.0`, because both now carry the `relations` and `storeys` blocks that 1.1.0 introduced (SPEC.md §20). A manifest using a 1.1.0 block while declaring 1.0.0 would be misdescribing itself. The field records which document the manifest was written against, not a difference in obligations — 1.1.0 is additive and a 1.0.0 document is still conformant input (§16.2) |
+
+Everything the gate reads is unchanged in kind: the normative `counts` block,
+with `source_ifc_type`, `status` (`known` / `known_gap` / `unsupported` /
+`decoder_baseline`), `tolerance`, `tracking_issue` and `unsupported_feature`
+carrying exactly the §6.1 meanings — joined since 1.1.0 by two blocks that
+work the same way. `relations` categories carry `relation_ifc_type` and
+`expected_pairs`; `storeys` categories carry `storey_ifc_type` and
+`expected_storeys`. Neither has a `tolerance`: their field class is exact set
+equality and §7.2 gives it no tolerance concept.
 
 The manifest additionally declares `semantic_surface` and `excluded` — the
-result that `counts` implies. `tools/verdict.py` derives both from `counts`
-and raises `MANIFEST_ERROR` if the declaration disagrees, so flipping a
-category's status without noticing is loud rather than a silent change to
-what is being claimed.
+result those three blocks imply. `tools/verdict.py` derives both and raises
+`MANIFEST_ERROR` if the declaration disagrees, so flipping a category's
+status without noticing is loud rather than a silent change to what is being
+claimed.
 
 ## What is in the corpus today
 
-Two artifacts, one edge, three witnesses.
+Two artifacts, one edge, three witnesses, three agreement classes.
 
 | Artifact | Alias | Bridge file | Verdict |
 |---|---|---|---|
-| `g-2026-0001` | magnetar-2024-core-interior | `2024_Core_Interior.ifc`, 20,392 bytes — an element-export fixture | PASS, 8 surface fields, 4 excluded |
-| `g-2026-0002` | magnetar-2024-core-interior-slim | `2024_Core_Interior_slim.ifc`, 1,665,968 bytes — the full project export | PASS, 4 surface fields, 9 excluded |
+| `g-2026-0001` | magnetar-2024-core-interior | `2024_Core_Interior.ifc`, 20,392 bytes — an element-export fixture | PASS, 6 surface fields, 10 excluded |
+| `g-2026-0002` | magnetar-2024-core-interior-slim | `2024_Core_Interior_slim.ifc`, 1,665,968 bytes — the full project export | PASS, 13 surface fields, 3 excluded |
 
 **They share the same `.rvt` and the same committed rvt-rs observation.** Only
 the Revit-authored bridge file differs, which makes the pair a direct measure
-of how much of the first artifact's agreement was real: the second has a
-narrower surface and five times the exclusions, because the full export shows
-360 walls, 132 doors, 256 columns, 116 spaces and 80 slabs where rvt-rs
-recovers 0, 0, 0, 18 and 64.
+of what a thin reference can support. The full export scores 360 walls, 132
+doors, 256 columns, 80 slabs, 20 shading devices and 15 storeys as exact
+matches, plus 138 `IfcRelFillsElement` host/filling pairs and the 15
+`[name, elevation]` storeys; the 20 KB fixture can score none of the element
+counts and no relation, but carries the same complete fifteen-storey
+hierarchy and so supports `storeys.IFCBUILDINGSTOREY` exactly as strongly.
 
 See the top-level [README](../README.md#what-exists-today) for what that is
 worth, and each artifact's PROVENANCE.md