track footnote and level support in api.features.xml

[webern]

Surfaced by the round-trip failure classifier (#211, PR #217).

What

footnote and level are each dropped in 373 failing files, but neither appears in data/api.features.xml. Because the classifier cross-references that index to decide whether a drop is "expected" (support="none") or a bug (support="full"/"partial"), files whose only blockers are untracked elements fall into the unknown bucket — the classifier can't categorize what it can't look up. This is a measurement blind spot, not an mx::api code bug.

What to do

Run the api-feature-audit skill to add footnote and level (and any other untracked-but-dropped elements it finds) to data/api.features.xml with the correct support level. Once tracked, the classifier can fold these files into B (drop-only, if support=none) or flag them as bugs (if support=full/partial), shrinking the unknown bucket.

Repro

make dump-api-roundtrip && make classify-api-roundtrip 2>warnings.txt
grep "footnote" warnings.txt   # "unclassified FAIL (missing=[... 'footnote' ...])"
grep 'name="footnote"' data/api.features.xml   # no match

References

• Part of api: expand corpus coverage of the api round-trip pass-list #208
• Found by api: classify corpus round-trip failures by root cause (drop-only, reorder, enum bug, missing feature) #211 / PR feat: dump and classify api round-trip failures #217

[webern]

Closing as resolved-by-removal rather than by tracking.

This issue assumed the fix was to add footnote/level to data/api.features.xml so the classifier could categorize the ~373 files that drop them. Investigating the editorial feature for #236 turned up the more fundamental fact: <footnote>/<level> appear in zero real-world corpus files. Every one of the ~382 editorial-bearing files was under data/synthetic/, where the audit generator stamps the boilerplate onto nearly every single-feature probe.

So the drops the classifier saw were noise, not a real api gap. Rather than model editorial in mx::api (the original #236 approach, now closed) or track a feature the wild corpus never uses, we removed the boilerplate at the source:

• feat: model editorial (footnote/level) in mx::api #236 — closed; editorial is intentionally not modeled in mx::api
• chore: strip editorial boilerplate from synthetic feature fixtures #239 — strips <footnote>/<level> from every synthetic fixture except the dedicated footnote/level probes

With the boilerplate gone, those ~373 files no longer drop footnote/level, so the classifier blind spot this issue described is gone — without adding the untracked-support entry that was proposed here. The dedicated footnote/level probes still exist for spec coverage, and if we ever want the audit to record them as support="none" that's a trivial follow-up, but it's no longer blocking the round-trip classifier.

---

Generated by Claude Code