v0.13.0

0.13.0 makes room and space import dependable on more files, reads Revit
2018 projects that exported nothing before, and shows the browser viewer's
real progress. Most of it follows reports from inevitable-co-ai (#1260,
#1261, #1262), hemanth1986 (#1075, #1247, #421) and puzzbobb (#421), and it
takes puzzbobb's speed-up of the native record reads (#1301). Thanks to all
three.

Measured against 0.12.2 on every reference model, Autodesk's sample projects
and the held-out models (Measure run 41741896192, this release's code
against v0.12.2):

• every IfcSpace says where its outline came from
  (RvtElementRecordGeometry.SpaceOutlineSource). On the BIM4LCA office 229
  of the 232 rooms paired with Revit's as built and 231 of 234 at building
  permit have Revit's area to 0.1% (220 and 221 with 0.12.2). Of the 12 rooms
  per file 0.12.2 drew as their box, 9 and 10 now have Revit's area; the rest
  are labelled approximate (a column Revit's export may not take off, 0.45%
  to 0.68% under Revit's floor area) or keep their box, and say so;
• rooms on Revit 2026 and 2027 take their outline instead of their box: each
  of the 257 rooms of Autodesk's four sample projects with rooms gives the
  same verdict and area in its 2026 and 2027 saves as in 2025, the rooms
  that keep their box in 2025 included; where an outline sits is not
  measured on those files;
• the BIM4LCA terrain office writes Revit's 9 IfcSpace, not 98: the 89
  bicycle bays Revit's export leaves out are left out;
• Revit 2018's ALD26 Before writes 1,136 of the 1,321 elements and all 73
  spaces of Revit's export, where 0.12.2 wrote none; its held-out sibling
  writes 1,336 elements and spaces against Revit's 1,520, none that Revit's
  export lacks (counts only). Revit 2016 to 2023 exports now leave out
  elements outside the export phase and in non-primary design options, as
  2024 and later did; on that one 2018 project the two filters together
  remove 198 elements Revit's export lacks, and the counts cannot say which
  removed which;
• on the BIM4LCA office 186 more columns as built and 250 more at building
  permit carry Revit's Length (279 of 279 and 321 of 327; the 6 at permit
  without it are not explained), and the two wall hatches have Revit's width
  (268 of the 268 doors Revit sizes). Both are one building's measurement;
• the T-shaped roof of puzzbobb's Revit 2025 20_T_U_Roof is Revit's body;
  the U-shaped roof stays its box, and roofs with several slope arrows are
  drawn on Revit 2025 only.

Faster, measured on one runner in one run against main before #1301 (which
already had #1237's 4% to 5%): #1301 makes Revit 2015 to 2023 exports of
Autodesk's samples 5% to 19% faster, their 2024 and 2025 samples 5% to 13%,
and the BIM4LCA office 7% to 9%; RE1 Architecture took 6% longer and RE1
Electrical 2% less.

What is not measured yet is said in each entry: among others, whether the
202 MB model behind #1075 now reads whole at the default limits, whether the
ElementIds of the #421 reporter's files are Revit's, the #1247 large-model
door fix on the user's own project, and the door width rule on any building
but the BIM4LCA office.

Upgrading from 0.12

• Every export's IFC header names rvt-rs 0.13.0, so an export that is
  otherwise unchanged still differs from 0.12.2's there.
• Every IfcSpace in a geometry export carries
  RvtElementRecordGeometry.SpaceOutlineSource (stored_solid,
  stored_solid_faces, stored_solid_approximate, record_box,
  saved_mesh, family_body or none). An importer can use it to tell a
  room drawn as its box from one drawn as its outline. The room schedule
  leaves the area of a boxed or approximate room blank.
• A family instance exported as IfcSpace whose family draws no model form
  for its parameters is no longer written, as Revit's export does not write
  it (#1262; measured on Revit 2024, applied to 2023 to 2027). The export
  diagnostics count it as element_record_family_space_without_body.
• Revit 2016 to 2023 exports leave out elements outside the export phase and
  in non-primary design options. Revit 2008 to 2023 projects whose element
  frames carry no ElementId of their own export many more elements (measured
  on one 2018 project).
• rvt-ifc writes the IFC to <output>.partial and renames it when whole,
  so a failed export leaves no truncated file under the output's name (with
  --observation it writes the file directly).
• At the default limits a partition may inflate to 16 times its stored size,
  1 GiB to 4 GiB on a 64-bit build, so --large-model is needed less often.
• A door or window sized from its symbol's box edges carries
  RvtElementRecordGeometry.SymbolWireWidth.
• rvt::ifc::step_writer::write_step_to streams the STEP text to any
  writer.
• The wasm binding openRvtBytesWithDiagnosticsModeAndOpenLimits takes an
  optional fourth argument, a progress callback.
• The export diagnostics add exported.bounding_box_spaces,
  exported.resolved_outline_spaces, exported.approximate_outline_spaces
  and a warning that counts doors and windows whose type record was not
  read.

Added

• The browser viewer shows the decoder's own progress, and says when a
  model is too large for a tab. The loading card names the stage the
  decoder is in (the stages rvt-gltf --progress prints) and fills its bar
  with that stage's count, where it used to sweep and guess "about N s" from
  the file's size. openRvtBytesWithDiagnosticsModeAndOpenLimits takes an
  optional fourth argument, a function called with the stage, the count done
  and the count it is out of. rvt-gltf --progress now also prints a
  partition-scan line for each Partitions/* stream as it is first
  inflated, which is most of a project's open, where it printed the stage's
  start and end only. A file whose size says it may need more than a tab's
  4 GiB (at least about ten times its size, the figure of the largest
  reference model; the warning can fire on a file that fits) is warned about
  before the decode and still tried. When the wasm instance runs out of
  memory, the viewer says the model is too large for the browser, how far the
  decode got and where, and that rvt-ifc and rvt-gltf with --large-model
  (streams up to 2 GiB and, on a 64-bit machine, inflating up to 8 GiB, for a
  file of up to 2 GiB), or Python with a larger max_stream_bytes and
  max_inflate_bytes, can read a larger file, instead of
  RuntimeError: unreachable. A stop a long way short of the ceiling says
  "Ran out of memory" and that the file may be damaged, and a panic that is not
  a memory failure keeps its own message (#1195).

• A column a clean join pairs with carries its Length. On a column
  drawn as its record box whose base quantities RE-291's rule writes,
  Qto_ColumnBaseQuantities.Length is now written as the box's height where a
  clean pair joins it: Revit's export writes it on every such column of the
  BIM4LCA office, 186 of 186 as built and 250 of 250 at building permit,
  equal to that height. The 18 columns Revit draws as a cut mesh, with no
  Length, are those a cut slab straddles, which the rule leaves out (#1203).

• Every IfcSpace's plan area and outline is scored against Revit's export,
  and says where its outline came from. Measure's new
  space_profiles_vs_ifc scorer pairs the spaces of rvt-rs's export with
  Revit's and counts those within 0.1% in area and 0.01 ft in outline (outline to outline, and by the mesh's vertices); the
  IfcOpenShell CI job runs it on Core Interior (116 of 116 rooms) and RE1
  Architecture (11 of 11) and fails below those counts. A room's
  RvtElementRecordGeometry.SpaceOutlineSource is stored_solid where its
  outline was read from the room's stored solid and record_box where the
  body is the record's bounding box and its plan area is the box's, not the
  space's (saved_mesh with --saved-meshes, family_body for a family
  instance or Revit Space drawn as more than its box, none for an Area,
  which has no body). A family instance written as a space gets the property
  too. exported.bounding_box_spaces in the diagnostics and a line in
  rvt-ifc's summary count the spaces drawn as their record's box. What an
  IfcSpace holds per release, and what is measured where, is in
  docs/rooms.md (#1260, RE-313).

Changed

• A partition may inflate to 16 times its stored size at the default
  limits, at least 1 GiB and at most 4 GiB on a 64-bit build. It was
  1 GiB whatever its size. A 202 MB Revit 2024 model whose 148 MB partition
  inflates past 1 GiB, which 0.6.1 read by cutting the partition short
  without a word, has needed --large-model since 0.11.0; at 16 times its
  stored size it should now read whole at the defaults, which is not yet
  measured on that model. A small file still cannot make a reader hold more
  than 1 GiB of one stream, and a 32-bit build and the browser viewer keep
  1 GiB (#1075).

• rvt-ifc writes the IFC as it builds it. The STEP text was built
  whole in memory and then written. It is now handed to the file every MiB,
  after the file handle and the partitions it keeps inflated are released.
  The bytes are the same as before this change, but for the version the
  header records. The text goes to a file beside the output, renamed
  when whole, so a failed export leaves no truncated file under its name. On
  the BIM4LCA office the text is 68 MB and the handle 362 MiB of inflated
  partitions; the saved-graphics pass, which runs before, still sets the
  peak there (#1075). With --observation the text is still built whole,
  since the witness reads it back. rvt::ifc::step_writer::write_step_to
  streams to any writer.

• Saved meshes read cone faces. A ConeSurf face failed with
  unsupported surface ConeSurf, which kept the whole element's box
  (RE-224). It is read as a cone of revolution about its apex, v the
  distance along the generator. With --saved-meshes RE1 Plumbing draws 8
  more pipe transitions, every Revit vertex within 0.001 ft of the mesh and
  all within 0.01 ft of Revit's body (1.08 to 1.10 of its volume: Revit's
  export is a 6- or 8-sided polyhedron). No other body scored against
  Revit's export changes, on RE1, the office or the Revit 2025 models, and no
  default export changes. On Autodesk's 2024 samples, which have no Revit
  export, --saved-meshes draws 471 more bodies: counted, not
  scored. The two RE1 Electrical panels #1144 names keep their box for
  trims RE-302 and RE-305 measured (#1144, RE-311).

• Exports read walls' native records once instead of twice, and skip
  per-record hashes no export uses. Measured on GitHub's runners against
  0.12.x on every reference model and Autodesk's samples, with output
  byte-identical but for the version its header records: Revit 2015 to 2023 exports take 4% to 5% less time, 2024 3%, 2025
  3%, 2026 and 2027 1% to 2%, the reference and BIM4LCA models 4%. Peak
  memory is unchanged within 1 MiB except on Autodesk's 2015 rme basic
  sample, 361 to 386 MiB, which is not explained (#1237).

• The export diagnostics count the doors and windows whose type record was
  not read. Since #1271, a door or window whose type's record a read did
  not reach (a large file over its limits) is written as a door or window
  drawn as its record box, not as its opening alone. Its category comes from
  its own record, but whether it draws geometry was never read. A warning
  now gives their number, since one that is a cased opening in Revit is its
  opening alone in Revit's export (#1247).

• A room whose stored solid the exact walk did not close is read again
  from the faces its solid lists, and labelled apart. The walk stopped at a
  record longer than 0x200 bytes between two faces, walls on lines up to
  0.0015 ft apart left a step no face crosses, a column overlapping a wall
  put a loop partly inside the wall's, and a solid ran past the 64 KB window.
  The solid ends at the count of its list of objects and at a gap of 1,024
  bytes, so the faces of the element that follows are not read, and the work
  is bounded (512 faces, 4,096 pieces; crafted solids take under 10 ms).
  SpaceOutlineSource is stored_solid_faces where the faces' edges close as
  the walk reads them, and stored_solid_approximate where a column overlaps
  a wall (its area depends on whether Revit's export takes the column off),
  an end of one wall edge lies on another, or the outline falls up to
  0.02 ft short of the record box; stored_solid
  stays the exact walk's. The diagnostics count both
  (exported.resolved_outline_spaces, exported.approximate_outline_spaces)
  and the room schedule leaves an approximate room's area blank. On the
  BIM4LCA office as built the 13 rooms that kept their box take an outline:
  nine are stored_solid_faces, all at Revit's area to 0.1% (their boxes were
  3% to 67% over it), and four stored_solid_approximate, 0.45% to 0.68%
  under Revit's floor area. At building permit 11 of 12 take an outline, all
  stored_solid_approximate, ten at Revit's floor area to 0.1%; room 3095831
  keeps its box. 6 of the 8 rooms that kept their box do on a held-out Revit
  2021 project, all 6 at Revit's area. Measure's space_profiles_vs_ifc
  compares the faces of Revit's mesh that point up, which read a room under a
  stepped or sloped ceiling as a fraction of itself (room 1671995: 1218 of
  5177 ft²; RE-313's 66% room); the new space_overlaps_vs_ifc also compares
  the floor, where all 443 office stored_solid room scores agree. The six
  rac_advanced rooms with round columns, 49 of the 50 rme rooms and
  Snowdon Towers' seven keep their box. Core Interior and RE1 Architecture are
  unchanged. A resolved room carries no NetVolume (#1260, RE-316).

• A roof sloped by several slope arrows is drawn as the planes they
  define, where Revit's export is measured to resolve them so. Each
  arrow's plane rises from the boundary line its tail is on, at the arrow's
  slope, with the top a plane's own thickness over the cosine of its angle
  above the underside, and the planes meet in hips and valleys. On
  puzzbobb's 20_T_U_Roof (Revit 2025) the T-shaped roof 4149 is now
  Revit's body, its farthest vertex 3.2e-06 ft from Revit's and its volume
  436.7078 ft³ against Revit's 436.7074 (it was the outline over the record
  box, 4,426.4 ft³). It is drawn on Revit 2025 only, for arrows set by
  their heights. The U-shaped roof 5941 stays its record box: Revit slopes
  its two long outer walls along the arms only, and a roof of its
  configuration (a sloped edge met inside by a reflex edge's extension,
  beside a sloped neighbour at a convex corner) is not drawn either, though
  only the U is measured (#885, RE-307).

• Native record reads are faster: the graph decoder borrows the schema's
  classes instead of copying them for every object, and a values read takes
  only the types its elements name. On one GitHub runner against main
  before it (Measure 41736361020), Revit 2015 to 2023 exports of Autodesk's
  samples take 5% to 19% less time (rac_advanced and rme 15% to 19%),
  their 2024 and 2025 samples 5% to 13%, and the BIM4LCA office 7% to 9%;
  RE1 Architecture, a one-second export, took 6% longer in that run and RE1
  Electrical 2% less. Peak memory is within 5% of before on every file, the
  largest rises Autodesk's 2015 rme_basic sample (361 to 379 MiB in that
  run) and RE1 Electrical (175 to 183 MiB). The IFC and the diagnostics are
  the same as the commit before this change (Measure 41740244025) (#1237,
  #1301; thanks to puzzbobb).

Fixed

• --large-model now raises the budget of the reads of native records,
  and a door or window is written as its opening alone only where its type's
  record was read. Each native read took a fixed 256 MiB budget for one
  reassembled group, which --large-model did not raise. On a user's 438 MB
  Revit 2024 project a group over it stopped the read of the door types and
  of the saved graphics, even with --large-model. That project's 73 doors
  then fell back to the type-parameter scan, which RE-299 showed misses
  maps, and were written as their openings alone (#1247). The default
  budget now follows the file's inflate limit, so --large-model raises it,
  and is never more than the partition's own inflated bytes. On Revit 2024
  to 2027, a door or window is written as its opening alone only where its
  type's native record was read with an empty geometry map; a type the read
  did not reach is not taken to draw nothing. Earlier releases keep their
  rule.

• Rooms on Revit 2026 and 2027 files take their outline, not their box.
  The reader of a room's stored solid ran on Revit 2008 to 2025 only, so on
  a 2026 or 2027 file every room was its record box, and an L-shaped room's
  box covered its neighbours: Autodesk's rac_basic Entry Hall was 72.8 m²
  where its 2025 save gives 30.3 m². Every room of Autodesk's four sample
  projects with rooms (257) now gives the same verdict and area in its 2026
  and 2027 saves as in 2025 (#1260, reported by inevitable-co-ai).

• A Revit 2008 to 2023 project whose element frames carry no ElementId of
  their own now exports those elements, on their storeys (measured on one
  Revit 2018 project). Revit 2018's
  ALD26 Before exported none of the 1,394 elements and spaces in Revit's own
  export (#1261): most of its frames in Partitions/0 do not open with a
  declared u32 ElementId, but sit inside their element's partition record, as
  2024 and later frames do (RE-35). Those records are u32 ElementId, u32
  size, the tag of ElementHeader, a count, the body and a 12-byte trailer that
  echoes the size; the 2024 walk's flag test stops at the first record with
  another flag word, after 115 of this file's 12,626. A placed-instance frame
  with no declared id now takes the id of the declared record around it when
  its own reference list names that id, and so does a Level frame that is a
  Level with a name block in its stream, which gives the file its 7 storeys.
  The file now writes 1,136 of Revit's 1,321 elements by ElementId, 1,127 of
  them in Revit's class or IFC4's for its IFC2X3 one (9 curtain walls come out
  as walls), none that Revit's export lacks, 1,073 on Revit's storey. A curtain
  panel frame that opens with a count rather than ff ff ff ff is left out,
  since 96 of the 102 on that file are the shells of panels Revit does not
  write. The 185 elements still missing are 178 conduits, conduit fittings and
  gutters, whose categories are not yet recovered, 4 panel walls and 3
  assemblies (RE-315). Files whose frames all open with their id export as
  before.

• Revit 2016 to 2023 exports leave out the elements Revit's own export
  leaves out by phase and by design option. Only 2024 and later applied the
  export phase (RE-184) and the primary design option (RE-40). A Revit 2018
  project wrote 198 pipes, fittings and other elements that Revit's export does
  not hold, and writes none now; the counts cannot say whether the phase or the
  design options removed them, and the held-out file is not opened to find
  out. The phase is read as RE-281 reads it. Design options are read as on 2024
  with u32 ids: on the rst_basic_sample_project of each of 2016 to 2023 the
  one Mass in a non-primary option is found and counted, as the 2024 and 2025
  copies of the model count it, and leaves the production counts of the
  diagnostics without changing the exported IFC. Revit 2008 to 2015 keep every
  element (RE-315).

• A family instance written as a space is left out where its family draws
  no model form for its parameters. Revit's export writes a family
  instance that is not a door or a window only when it has a 3D body. On the
  BIM4LCA terrain office, 92 of the 94 "Parking Space" instances whose
  "Export to IFC As" is IfcSpace are bicycle bays whose "Stripe Y/N" is 0,
  so no form of their family is drawn, and Revit's export holds only the two
  car bays; rvt wrote 89 of the bicycle bays as IfcSpace too. The terrain
  office's IfcSpace are now the 9 of Revit's export (9 of 98 before), and
  the six reference models, the other office files, the R2025 and the
  held-out models export byte for byte as the commit before this change
  did. Only instances written as
  IfcSpace are tested, other categories are not, and the export
  diagnostics count what is left out (#1262).

• A 28-byte Global/ElemTable is recognised by its first id, the one
  declared_ids reads as the ElementId. The layout check tested the
  second id, which on 8 client files saved by Revit 2024 to 2026 with a 2022
  or 2023 schema holds an older, lower number on 1% to 37% of records, so
  their tables were read as 12-byte records with a "fits neither its 28- nor
  its 40-byte layout" warning. On the reporter's build of this change the
  warning is gone on all 8; whether their records' ElementIds are Revit's is
  not yet measured (#421, found by puzzbobb, measured by hemanth1986).

• A door or window whose symbol cuts no loop takes its OverallWidth
  from the lines of no category in that symbol's saved graphics, and the
  BIM4LCA office's two wall hatches now have Revit's (RE-318, #810).
  Doors 1756089 and 1765267 are TRAPDOOR hatches (family LU-3.VSS, types
  _90° and _180°), not doors with a sidelight. Their record box was
  0.1137 ft wider than Revit's 5.8629 ft, and neither their cut loop (none),
  cut solid, stored lengths nor visible surfaces give it: the 90° type hides
  the 180° leaf its width includes. A symbol's saved graphics hold the edges
  of the boxes that bound its forms as lines of category -1, hidden forms
  among them, and their extent along X is Revit's width on all 268 of the
  office's doors Revit sizes and all 269 windows as built (268 and 265 at
  building permit), where the record box missed 2 doors and the surfaces 2.
  The rule fires only where the record box sized the door or window before,
  and a symbol with no such line, a degenerate extent, an instance or
  component reference in its graphics, two readings that disagree, or edges
  reaching past the record box, keeps the width it had. Against main
  (Measure 41725344128) the office's IfcDoor widths equal Revit's on 268
  of 269, from 266, and three widths
  move on the as-built file (1756089, 1765267 and, by 7 micrometres, 1722215);
  the only other model with such a door, RE1 Architecture's curtain-wall
  door 445975, keeps its width, and 31 of 34 measured models and all 18
  held-out models write the same IFC as the commit before this change. One
  building's families: what the
  boxes are, and why their extent is Revit's width, is inferred from Revit's
  widths, not documented (reports/element-framing/RE-318-symbol-wire-width.md).

Research

• Why saved pipe fittings measure 1.2 times Revit's volume (RE-311).
  Revit's export of a pipe elbow is a polyhedron of 6, 8 or 10 sides round
  the pipe and two segments along its bend, inscribed in the surface: 16 to
  26% short of the true torus, which rvt-rs's mesh reaches to within 1 to 2%.
  The closet bends' wrong volumes (0.28 to 2.59, one negative) were the
  scorer's, which took a mesh with cracks between its faces about the file's
  origin; mesh_bodies_vs_ifc.py now takes it about the body's centre (the
  three RE1 Mechanical bends: 1.1431 each) and counts the bodies whose Revit
  vertices all lie on the mesh. fitting_faces_vs_ifc.py and
  probe_re311_fitting_faces set each face of a saved fitting against
  Revit's body (#1144).

• Measure scores the BIM4LCA terrain office, in concrete and in timber,
  against Revit's exports of them. -f nordic=true now reads the two
  files at building permit and their IFC4 exports by byte range, beside the
  concrete office. probe_re312_parking_spaces and
  tools/re/parking_spaces_vs_ifc.py list the family instances written as
  IfcSpace against Revit's export and the model forms of their families
  (RE-312, #1262).

User guide: https://github.com/DrunkOnJava/rvt-rs/blob/v0.13.0/docs/user-guide.md