[DrunkOnJava]

Summary

Draft. #1498's structural fix and the additive sources for the inputs it lists. Refs #1498 (not Closes until the test has run green on the gate and the proof run is read).

• property_readers (src/ifc/element_provenance.rs) lists a noted source for any reader an element's own *Source property values name (BodyHeightSource, BodyBaseSource, BodyEndSource, PlacementAxesSource, ThicknessSource, SpaceOutlineSource), beside its BodySource and ProfileSource. Before, only those two keyed a listing, so a door whose body was its saved mesh never listed the transform its placement was taken from.
• SOURCE_VALUES in tests/value_sources_offsets.rs gives every written *Source value a row (sourced, covered by another table, unread, or a declared gap with its reason) and fails on a value with none. A sourced row requires its readers on every element that carries the value. It reads the real exports of the CI gate through --element-provenance.
• New sources, all provenance only and noted only while the table is being built (recording_value_sources()): native_wall (the wall's own native record: join lists, cut pairs, m_realJoined, the curve, the flip), native_wall_joined (the records of the walls the lists name, where every one was read), wall_top_covers and wall_base_cover (the frame of each cover the wall names, rule frame), wall_orientation (agreeing, the scan compares every copy), wall_type_face_angle (agreeing), native_opening (a door's own record: the two flips), opening_host_wall (the host wall's record).

Declared gaps: 20 values

12 bindings to a Level or a storey (LevelBindSource 9, StoreyBindSource 3), 2 family names chosen by a table (FamilyNameSource), 1 column height (BodyHeightSource partition_geometry_node_after_joins: the geometry node's offset is not noted), 5 schema-walker values no model of the 92 writes. Reasons and element counts: docs/export-diagnostics.md, "Written *Source values". Three listed inputs have no *Source property to key a row on and carry a stated reason instead: door and window cut-loop sizes, a beam's cross-section rotation, a section's dimensions (not compared with the export).

Evidence so far

• Measure run 42136996423 (probe at 92a7e92, base=none, 92 models, scorers=none): the inventory of every *Source value and the readers behind it. It found three values that had no row (StoreyBindSource record_level_reference, LevelBindSource structural_framing_top_level, ProfileSource partition_element_record_join_cut); they have rows now. That run is the probe commit, which has no export code: it says nothing about byte identity.
• Byte identity of the export is not yet measured on this head. The one filtered run (-f base=main -f samples=true -f sample_years="2024 2027" -f models="core-interior nordic-office-*" -f nordic=true -f scorers=none) waits for the second read.
• CI on this head: not yet read.

MCP contract

No rvt-mcp file is touched. The value_sources description (rvt-mcp/src/schema.rs, docs/mcp-tools.json) says the copy each reader took "a value of the element's body from"; with this PR it also lists the values of the wall rules and the placement. The wording is the coordinator's to sync; the reader names are free strings in the schema.

Testing

• cargo fmt, tools/release/changelog.py check, tools/ci/api-doc-ids.py locally.
• tests/value_sources_offsets.rs on the gate in CI (Corpus health): not yet read at this head.

Breaking changes

None.