The two ends of one report disagreed about the same files

[melbinjp]

The header asks likeliest_docs_directory whether anything unread is NAMED like documentation, and prints either a widen suggestion or "none of them is named like a documentation tree". The verdict at the bottom then called every one of those files "documentation file(s)" and divided by them.

On zhukunpenglinyutong/desktop-cc-gui the same run says, forty lines apart:

   5812 documentation file(s) elsewhere in the tree were NOT read
     openspec/ 5271, src-tauri/ 251, .agents/ 232, dev-guidelines/ 45, ...
     none of them is named like a documentation tree
...
This judged 63 of 5875 documentation file(s). 5812 were never read, so this
verdict covers 1% of the documentation in this project.

5,271 of those are an OpenSpec change-proposal tree. Nobody would want them checked against HEAD, the tool has already said as much, and then it reports 1% coverage of "the documentation in this project".

Measured, because one repository is not a defect

39 sweep captures, 24 of which print a coverage line:

coverage range	1% to 99%, median 27.5%
under 10%	7 of 24
a docs-named directory found unread	4 of 24

So in twenty of twenty-four cases the tool divided by a denominator it had itself just judged to contain no documentation tree.

The fourteen commonest biggest-unread directories, and exactly one is documentation:

skills/ 2   src/ 2   .agents/ 2   scripts/ 1   ts/ 1   datafusion/ 1   crates/ 1
libs/ 1   benchmarks/ 1   examples/ 1   docs.feldera.com/ 1   tools/ 1
mobile/ 1   mcpjam-inspector/ 1

Per-package READMEs and agent instruction files, which find_docs excludes on purpose and whose docstring argues why.

What changes and what does not

The percentage stays. A narrow pass must never read as a clean one, which is the entire reason this line exists - the README promises "a clean report over two files in a project with three hundred cannot be mistaken for a clean report over three hundred" and printing it forty lines from the verdict is how it got mistaken in the first place.

What goes is the assertion about what those files are:

- This judged 63 of 5875 documentation file(s). 5812 were never read, so this
- verdict covers 1% of the documentation in this project.
+ This judged 63 of 5875 documentation file(s) in the tree, 1%. The 5812 it did
+ not read are outside the default scope and listed above; none of them is in a
+ directory named like documentation.

And the other branch, which is the one the change exists for - when something unread really is a documentation tree, the last line says so rather than making the reader scroll:

This judged 5 of 219 documentation file(s) in the tree, 2%. The 214 it did not
read include `docs.feldera.com/`, which is named like a documentation tree, so
this verdict is narrower than it looks.

report_coverage now returns the directory it computed instead of throwing it away, and the verdict carries the same judgement the header already made.

209 tests. Two rewritten to the new wording, one new for the docs-tree branch, and the new one also asserts the header still prints the actionable --docs form, so the two ends are checked together.