[elvogel]

The audit that produces docmd's published coverage figures measured less than the product
supports, and the figures it produced were a year-old release behind. Both fixed here, in that
order, because the second depends on the first.

The instrument

Corpus_Audit globbed *.docx. The converter accepts .docx, .docm, .dotx and .dotm, and
on Linux that glob matches case-sensitively as well -- so a .dotx or a .DOCX in an
audited folder was left out of the figure rather than counted. One of the 50 documents sampled
for this measurement is a .dotx, and the tree it came from holds seven files with an upper-case
extension.

private static readonly string[] WordExtensions = [".docx", ".docm", ".dotx", ".dotm"];

private static bool IsWordDocument(string path) =>
    !Path.GetFileName(path).StartsWith("~$", StringComparison.Ordinal)
    && WordExtensions.Contains(Path.GetExtension(path), StringComparer.OrdinalIgnoreCase);

~$ is Word's own lock file -- ownership metadata, never a readable package -- so it is skipped
rather than counted as a failure.

A file that is not an OPC package at all is now its own reported category instead of failing
the audit's one hard assertion. A real folder of client documents usually holds one: a Word
97-2003 binary that someone renamed to .docx. The CLI already calls that input error and exits
2 rather than 1, and the audit now draws the same line. Every other exception type still fails
the run, because an exception really does end a batch.

The figures

Re-measured on 0.3.0, corpus 394b2cb31e7459a1, 2026-10-15, run time 13 m 44 s:

Word documents found	50
Not a Word package	1 — reported, and excluded from the denominator
Converted without error	49 of 49
Lost not one word	45
Words lost	21 of 245,034 — 0.009%
Worst document, by words	11 words (99.8% kept)
Worst document, by share	2 of 177 words (98.9% kept)

Every one of the 21 words falls into a cause docs/limitations.md already records -- unmapped
w:sym characters, <pict>/<shape> wrappers, words fused at a figure-caption boundary, one
bare URL -- and the two causes 0.3.0 fixed produced no losses at all.

The corpus the old figures used is unreachable

README.md and docs/limitations.md quoted 27 of 50 against dca5084324a9022c. That corpus
cannot be re-measured. A corpus id is a hash of its documents' content, so it recognises a
corpus and cannot reconstruct one, and the _source.tsv that mapped those documents was not
kept anywhere on the machine.

This is the second time a corpus behind a published figure has become unreachable, and the first
time is already written into scripts/verify-corpus.sh: "by the time anyone tried to re-measure,
20 of its 21 names no longer resolved to a file." Content hashing was the fix for names; keeping
the manifest is the fix for hashes. Both documents now say so.

The two figures are therefore not comparable -- different documents and a different
converter. Both documents say that too, rather than inviting the reader to read an improvement
into the difference.

Also

Two limitations still said Status: open for defects that shipped in 0.3.0 -- the tab-indented
code block (#38) and the non-numeric w:ilvl (#37). Both are marked fixed, with the commit and
the verification that closed them, in the style the file already uses for fixed entries.

Suite: 394 total, 393 pass, 1 skip -- the skip being this audit with no DOCMD_CORPUS set,
which is how it is meant to behave.

🤖 Generated with Claude Code

https://claude.ai/code/session_018wZEgtuzaZPswykiGEBxbz