fix(changesets): correct twelve semver bumps that would have shipped breaking changes as patch (#3175)

The release was one command from publishing. A changeset decides the version
BUMP as well as the changelog text, and nobody had reviewed the changesets --
CodeRabbit was rate-limited across the batch and both review passes were aimed
at code and tests.

Twelve frontmatter corrections. Each was derived from what the referenced
commit actually did to the public surface, not from the changeset's own prose,
and the level follows AGENTS.md:68 ("major (>=1.0 pkg) or minor (0.x), never
patch when the surface shrank") rather than generic semver:

  bcf-validate-against-official-xsds       bcf patch -> major
      writeBCF throws where it returned a Blob. Probed both sides: at e19aa0e^
      "wrote blob, size = 1662"; on main "THREW -> BCF 3.0 requires a positive
      Camera/AspectRatio". No camera factory in the package sets that field, so
      the package's own documented createBCFFromIDSReport -> writeBCF path fails.
  recommended-scale-units                  drawing-2d + sdk minor -> major
      getRecommendedScale throws for 0/negative/NaN/Infinity where it returned
      1:1000, and the SDK forwards with no catch.
  export-step-fails-closed-on-...          wasm patch -> major
      export_step_json went String -> Result<String, String> on a root-level
      pub use. The Rust crate version derives from the highest npm workspace
      version, so patch would publish ifc-lite-export 5.0.1 with a hard compile
      break inside ^5.0.0. Precedent: #2766 took major for this shape.
  ifcx-writer-discards-globalid            ifcx patch -> major
  ids-summarize-not-applicable             sdk patch -> minor, + cli patch
      New required property on a RETURNED type. Minor, not major: nothing in
      the SDK accepts an IDSValidationSummary, so consumers never construct one.
  branch-merge-layer-overlay               collab patch -> minor
  collab-snapshot-invented-provenance      collab + ifcx patch -> minor
  symbolic-cycle-truncation-reason         server-client patch -> minor
  symbolic-output-cap                      + wasm minor
  csv-numeric-cells-export-as-numbers      + wasm major
  csv-cell-escaper-unification             + viewer patch
  sun-shadows-depth-prepass                + viewer minor

Effect: packages receiving a major bump go from four to eight. bcf, ifcx, sdk
and wasm would each have shipped a breaking change as patch or minor.

A fourth body correction, found by review of this commit rather than of the
batch: changing recommended-scale-units to major left its own body still
arguing "this is minor rather than patch". Correcting a bump without correcting
the sentence that justifies it reproduces the exact defect this commit exists to
fix. Rationale rewritten to state the narrowed input domain as the reason.
Swept the other eleven for the same contradiction; none.

Three body corrections, where the text ships AS the changelog:

  wasm-step-export-non-ascii-x2-directive  said encode_ifc_string "was never
      wired into this writer; it now is". It is still not called: the function
      is defined at rust/core/src/step_encoding.rs:146 and its only appearance
      outside rust/core is a comment. The encoding was reimplemented inline.
  bounded-glb-rep-instancing               announced export_step_to_writer under
      a wasm heading. It is exported from the Rust crate and appears nowhere in
      rust/wasm-bindings, so npm consumers cannot reach it. Qualified.
  wasm-bound-symbolic-item-walk-cycle      said "none of the three bounds
      reports anything". That was true when written; a later change in this same
      release gives them a truncation reason, so the shipped changelog would
      have contradicted itself. The stale sentence is DELETED rather than
      annotated: a first attempt added a cross-reference to the superseding
      changeset, but changesets never render the changeset filename, so it
      pointed at something no reader can find -- and the two entries land in
      different CHANGELOG sections (patch vs minor) anyway. The superseding
      entry states the current behaviour positively, so nothing is lost.

NOT changed, after two reviewers disagreed and the evidence settled it:

  placement-memo-truncation, geometry-bound-layer-slicing-identity-chase and
  validate-export-formats-externally were flagged for naming an npm package
  against a Rust-only commit. That is the documented house convention, stated in
  74a55a9: "Attribution follows the split precedent already on main:
  rust/geometry -> @ifc-lite/geometry". Correct as they stand.

  csv-numeric-cells-export-as-numbers was flagged as export minor -> major for a
  flipped default. escapeCsvCell and guardSpreadsheetFormula have never been
  published: added by #3102, still unreleased, shipping for the first time in
  this release. npm view @ifc-lite/export version = 2.9.4 and its CHANGELOG has
  no csv-cell entry. A default cannot break callers of a function no consumer
  can have called. Minor is right; only the missing wasm entry was real.

A fifth body correction: dxf-insunits-survey-units stated the scale error
BACKWARDS. It said a survey-foot DXF "came in at roughly 1/3.28 scale", i.e. too
small. `unitScale` is a MULTIPLIER (convert.ts:232 `wx = p.x * state.unitScale`)
and the unknown-code fallback is `unitScale = 1` (convert.ts:129, warning
"drawing units treated as metres"), so coordinate 100 in survey feet became
100.0000 m against a true 30.4801 m -- 3.28x TOO LARGE. Corrected.

Two reviewers had independently failed to find that import path and this was
recorded as unverified rather than edited on a guess; a third look found it at
packages/drawing-2d/src/dxf/convert.ts:48. Recording the gap rather than
guessing is what made it findable.

Author: louistrue

.changeset/bcf-validate-against-official-xsds.md

@@ -1,5 +1,5 @@
 ---
-"@ifc-lite/bcf": patch
+"@ifc-lite/bcf": major
 ---
 
 Validate real `.bcfzip` output against buildingSMART's published BCF XSDs. That check found eleven schema violations in our output; nine are fixed here and two are reported below as needing a decision this change cannot make.

.changeset/bounded-glb-rep-instancing.md

@@ -20,6 +20,6 @@ Grouping follows `collate_refs` exactly: a rep identity whose occurrences disagr
 
 f32 output only. A quantized shared mesh carries a non-uniform dequant scale that cannot fold into a rotating placement without breaking `Matrix4.decompose`, so quantized output still goes out flat.
 
-Also adds `export_step_to_writer`, which emits each record as it is read instead of returning the whole file as a `String`. On a 1 GB model that removes a gigabyte of output from the peak. It buffers internally, so a caller passing a bare `File` does not pay two syscalls per record.
+Also adds `export_step_to_writer` to the `ifc-lite-export` Rust crate (it is not exposed through the wasm bindings, so npm consumers cannot call it), which emits each record as it is read instead of returning the whole file as a `String`. On a 1 GB model that removes a gigabyte of output from the peak. It buffers internally, so a caller passing a bare `File` does not pay two syscalls per record.
 
 Output is byte-identical for models with no shareable shapes.

.changeset/branch-merge-layer-overlay.md

@@ -1,5 +1,5 @@
 ---
-'@ifc-lite/collab': patch
+'@ifc-lite/collab': minor
 ---
 
 Fix `mergeBranch(parent, branch, 'layer')` silently dropping every edit the branch made to an entity that already existed in the parent.

.changeset/collab-snapshot-invented-provenance.md

@@ -1,6 +1,6 @@
 ---
-'@ifc-lite/collab': patch
-'@ifc-lite/ifcx': patch
+'@ifc-lite/collab': minor
+'@ifc-lite/ifcx': minor
 ---
 
 Stop a collab snapshot round trip from inventing per-entity provenance, and carry the real thing on the wire.

.changeset/csv-cell-escaper-unification.md

@@ -3,6 +3,8 @@
 "@ifc-lite/sdk": patch
 "@ifc-lite/cli": patch
 "@ifc-lite/mcp": patch
+"@ifc-lite/viewer": patch
+"@ifc-lite/wasm": patch
 ---
 
 CSV cell escaping now has one implementation per language

.changeset/csv-numeric-cells-export-as-numbers.md

@@ -1,11 +1,12 @@
 ---
 "@ifc-lite/encoding": minor
 "@ifc-lite/export": minor
-"@ifc-lite/lists": patch
+"@ifc-lite/lists": major
 "@ifc-lite/viewer": patch
-"@ifc-lite/sdk": patch
-"@ifc-lite/cli": patch
-"@ifc-lite/mcp": patch
+"@ifc-lite/sdk": major
+"@ifc-lite/cli": minor
+"@ifc-lite/mcp": minor
+"@ifc-lite/wasm": major
 ---
 
 CSV: numeric cells export as numbers. **The formula guard's default changed.**

.changeset/dxf-insunits-survey-units.md

@@ -4,4 +4,4 @@
 
 Fix DXF import treating $INSUNITS codes 17-24 (gigametres, astronomical units, light years, parsecs, and US Survey feet/inch/yard/mile) as unknown and falling back to metres.
 
-The most consequential of these is 21 (US Survey Feet): a civil/survey DXF authored in that unit previously came in at roughly 1/3.28 scale — a large, visible error — even though the fallback correctly warned about the unknown code. US Survey Feet is now converted using its exact legal definition, 1200/3937 m (≈0.3048006096012192 m), not the international foot (0.3048 m exactly); the ~2 parts-per-million difference between the two is the entire reason US Survey Feet is a distinct $INSUNITS code, so Survey inch/yard/mile are derived from the survey foot rather than from international units.
+The most consequential of these is 21 (US Survey Feet): a civil/survey DXF authored in that unit previously came in at roughly 3.28x its true size — a large, visible error — even though the fallback correctly warned about the unknown code. US Survey Feet is now converted using its exact legal definition, 1200/3937 m (≈0.3048006096012192 m), not the international foot (0.3048 m exactly); the ~2 parts-per-million difference between the two is the entire reason US Survey Feet is a distinct $INSUNITS code, so Survey inch/yard/mile are derived from the survey foot rather than from international units.

.changeset/export-step-fails-closed-on-malformed-mutations.md

@@ -1,5 +1,5 @@
 ---
-'@ifc-lite/wasm': patch
+'@ifc-lite/wasm': major
 ---
 
 Fix `exportStep` silently discarding the caller's edits when `mutationsJson`

.changeset/ids-summarize-not-applicable.md

@@ -1,5 +1,6 @@
 ---
-"@ifc-lite/sdk": patch
+"@ifc-lite/sdk": minor
+"@ifc-lite/cli": patch
 ---
 
 Fix `bim.ids.summarize()` counting a `not_applicable` specification as passed. A specification whose applicability matches zero entities and whose cardinality does not require a match (no `minOccurs`) is neither a pass nor a fail — `@ifc-lite/ids`'s own `validateIDS` report already treats it that way — but `summarize()` had no `not_applicable` bucket, so its unconditional `else` folded every such specification into `passedSpecifications`. That inflated the spec-level pass rate returned by the CLI's `ids --json` output relative to the CLI's own text-mode output (both should read from the same validation, but text mode reads `report.summary` directly while `--json` goes through `summarize()`).

.changeset/ifcx-writer-discards-globalid.md

@@ -1,5 +1,5 @@
 ---
-"@ifc-lite/ifcx": patch
+"@ifc-lite/ifcx": major
 ---
 
 Fix `IfcxWriter` discarding every entity's IFC GlobalId.