[DrunkOnJava]

Summary

Revit 2008 to 2023 element frames that carry no ElementId of their own (the second prologue, RE-30) now take the id of the declared partition record around them, as 2024 does (RE-35), and so do Level frames that are a Level with a name block. Measured on one Revit 2018 project; no other release of that layout is measured for it. Revit 2018's ALD26 Before goes from 0 elements to 1,136 of Revit's 1,321, 1,127 of them in Revit's class or IFC4's (9 curtain walls come out as walls), none written that Revit's export lacks, and 7 storeys. On Revit 2016 to 2023 the export phase and non-primary design options are applied now, as 2024 does, which removed the 198 extra elements the held-out ALD26 Project wrote.

Motivation

#1261: a Revit 2018 project exported no building element against Revit's own 1,394, because most of its frames do not open with a declared u32 ElementId, and the 32-bit reader only read frames that do.

Changes

• src/research/partition_element_records_2023.rs: record_spans walks a partition's records (u32 id, u32 size, ElementHeader tag, count, body, 12-byte trailer echoing the size) on the tag and the echo alone, where the 2024 flag-word test stops at the first record with another flag word, and resumes at the next record that validates (one memmem Finder). scan_records gives a placed-instance frame with no declared id the id of the declared record around it when the frame's reference list names that id. A curtain panel frame whose first word is a count gets none (96 of the 102 on the file are shells of panels Revit does not write). EXPORT_FILTER_RELEASES (2016 to 2023) gates the filters below.
• src/research/partition_level_records.rs: a standalone placed OST_Levels frame with no declared id takes its enclosing record's id. A Level frame's reference list is empty and its name block lies elsewhere in the stream, so the guard is its class word being the schema's Level tag and a name and elevation block of that id in the stream (a stray Level-category frame cannot add an id with no name and drop the storeys). A frame that opens with its id is read as before.
• src/research/partition_schema_mvp/records_2023.rs, partition_design_options.rs, native_phases.rs: on 2016 to 2023 a selected instance that PhaseFilter::excludes or DesignOptions::excludes is dropped (rooms never by phase); design options are read with u32 ids (set record, name entry, the u32 26 bytes before a frame's marker); the diagnostics count both as on 2024. 2008 to 2015 read no set and keep every element.
• src/ifc/mod.rs: the export warning says per release what is applied (the export phase and design options on 2016 to 2023, neither on 2008 to 2015) instead of that design options are not read for every 2008 to 2023 export. partition_design_options.rs searches a file's option sets for 64 at most (OPTION_SETS_MAX): the rest stay unresolved and keep their elements, as a set with no name entry does on 2024; the most any file here has is 2.
• probes/examples/probe_re315_32bit_record_chain.rs, reports/element-framing/RE-315-32bit-second-prologue-records.md; tools/re/instances_vs_ifc_tags.py accepts the --list Measure passes a probe scorer; docs, support matrix and a changelog fragment.

Testing

Unit tests are banned in this repository; the evidence is Measure on real files.

• CI at a3d1cd2a: CI passed and 26 other checks pass; the viewer build (not required) was still pending at the last poll
• Verified end to end on real Revit files, against Revit's own export. Measured head 01b5af37 (the code of the figures below; the later commits change the export warning's text, cap the option sets searched, merge main and change documents, and the final-head runs repeat the byte identity):
  • ALD26 Before (Revit 2018, CC BY 4.0), counts run 41707934595 and probe run 41707939060: main writes 0 elements; the branch writes 1,209 building elements and spaces against Revit's 1,394. 1,136 of Revit's 1,321 elements are written and paired by Tag, 1,127 of them in Revit's class or IFC4's for its IFC2X3 one; 9 IfcCurtainWall are written as IfcWall. None written that Revit's export lacks; 7 storeys, as many as Revit's, and 1,073 of 1,136 elements on Revit's storey. Missing 185: 178 conduits, conduit fittings and gutters (found and attributed to Revit's exact Tags by the probe, not written: categories -2008132, -2008128, -2001391 are not in RECOVERED_CATEGORIES), 4 panel walls, 3 assemblies. Without the Level change (run 41707946855, commit 9ff5e950) 687 of the 1,136 are on no storey.
  • Level guard: the schema's Level tag and the class word of the Level frames agree on Before (2130: 7 standalone, 16 container members; 2131 the type symbol) and on the standalone Level frames of rst_basic_sample_project 2018 to 2023 (2130, 2164, 2198, 2199, 2299, 2377; 54 of 54); all 7 of Before's survive the check.
  • Held-out ALD26 Project, counts only (same run 41707934595): main writes 0; the first reading of the branch, before the filters (run 41713761592, commit 8fc583a1), wrote 1,534 elements and spaces, 198 of them elements Revit's export lacks (95 pipe segments, 90 pipe fittings, 8 proxies, 2 roofs, 2 windows, 1 wall); with the filters it writes 1,336 (1,263 elements and 73 spaces) against Revit's 1,520, no element Revit's export lacks, 1,255 of its 1,263 elements in Revit's class or IFC4's (8 curtain walls as walls), 1,181 on Revit's storey. The counts cannot tell which filter removed the 198.
  • Evidence for the filters. Phase: only Project's counts; Before and all 33 Autodesk samples of 2015 to 2023 have no element outside their last phase. Design options: eight saves of one model (rst_basic_sample_project, 2016 to 2023), 1 set and 4 frames in a non-primary option each, a Mass Revit does not export, so no export effect is observed; the exported IFC of all eight is byte for byte main's and only the diagnostics change (a skipped item with 1 Mass, the Mass leaves production_class_counts, production_walker_elements falls by 1).
  • Frigento, Gdynia, Warmond and the held-out BIM Whale AdvancedProject: IFC byte for byte main's, BIM Whale's counts lines main's. Frigento (22) and Gdynia (3) still write elements Revit's export lacks, from another cause.
  • Byte identity, run 41707942875 (-f samples=true -f nordic=true -f r2025=true): 92 models, 0 differ from main's IFC, the 33 Autodesk samples of 2015 to 2023 among them; the only diagnostics that change are the eight rst_basic saves of 2016 to 2023
  • Final head 68ae7904 (the merge of main 53e06c02 into the code): unfiltered held-out run 41722201615 (-f holdout=true -f base=main, counts only for the held-out): the IFC of 23 of 25 models is main's; the two that move are the two Vitoria saves, and the counts.txt of the other 17 held-out models, modelo_bim (2023) among them, is main's line for line; identity run 41722205132: 92 models, 0 differ from main's IFC (the diagnostics change on the 33 samples of 2015 to 2023, the warning's text, and the skipped items on the eight rst_basic saves).
  • Reproduce: the commands in RE-315 §8.
• If this touches the IFC exporter: witness verdicts replay: no Core Interior export moved (Corpus health and the IfcOpenShell job pass)
• If this is a performance change: N/A

Audit-honesty checklist

• README / docs reflect the current state after this PR: docs/evidence.md, docs/supported-profile.md, docs/support-matrix.json (still experimental), docs/reference-models.md, docs/export-diagnostics.md, docs/status.md, docs/format.md, docs/rooms.md, README.md, ROADMAP.md and W2 of docs/release-1.0-criteria.md (two projects now have a same-release export, 2015 and 2018, and every statement that said one now says two)
• Changelog fragment changelog.d/fixed/1261-32bit-second-prologue.md
• Limitations stated here, in the report (§9, §10) and in the code: only Revit 2018's ALD26 Before is measured against Revit's export for this kind of frame; conduits, conduit fittings and gutters not exported; panel shells and 4 panel walls; 3 assemblies; 9 curtain walls written as walls; the phase filter's only evidence is the held-out counts; design options are measured on Autodesk's rst_basic samples only, with no Revit export of them; 2008 to 2015 apply neither filter
• Conventional Commits, no emojis, no private paths
• Reverse-engineering finding: probe under probes/examples/ and report RE-315

Legal and contribution hygiene

• Apache-2.0. Clean-room: only public byte observations of a CC BY 4.0 model, Autodesk's published sample projects and Revit's own IFC exports; no Autodesk SDK, disassembly or NDA material.

[DrunkOnJava]

Automated Claude Code review (the maintainer's review agent; it read the code, the report and the Measure runs, over two rounds): three blocking items in the first round, fixed in 01b5af3, b036c93 and 2869ecd. The status pages still called the second-prologue ids and design options 2024 and 2025 only; the class claim said every element was in Revit's class where 9 curtain walls are written as IfcWall (1,127 of 1,136); and the report quoted counts from a held-out model beyond its counts lines. The filters were also limited to Revit 2016 to 2023, and a Level frame takes an enclosing id only when its class is the file's own Level class. Two blocking items in the second round, fixed in c04c92d: five pages still said only one 2015 project has its release's own export, and every 2008 to 2023 export warned that design options are not read, which is no longer true on 2016 to 2023.