OI-43 step 4: say which languages a repo could only partly read

[BrettCrawley]

OI-43 step 4 — say it in the output, not only in a changelog

The gate records the matrix and the changelog explains it, but neither reaches
someone reading a repo's findings. A Go or TypeScript repo looked exactly like a
Java one: findings present, confidence stated, and nothing to say that T1 and T2
resolution could not fire at all for it.

go: limited extraction across 40 file(s) — declared field types, so T1
resolution cannot fire; supertypes, so T2 resolution cannot fire. Findings for
this language are incomplete rather than absent; see OI-43.

One note per repo per language, never per file. Scala alone would otherwise
put a note on every Scala file in an estate, and a signal that loud stops being
read — the noise question that kept this out of OI-36 phase 1. A Java or Kotlin
repo stays quiet, because those are the covered ones, and a signal that fires
on success is not a signal.

The gaps are computed, never restated

A hand-maintained list of "languages we do not fully support" is exactly what
rotted into OI-43, and it would go stale the moment a table gained an entry.
coverage_gaps() reads the live grammar tables, so a note cannot claim a
limitation that has since been fixed, nor stay silent about one that has not.

test_a_filled_table_silences_its_gap fills a table and asserts the gap
disappears — proving the derivation is real rather than incidental.

One pass, not two

run-manifest.json gains counts.resolution_gaps — repos affected per language
— computed in the same pass as counts.unparsed. Reading the fleet twice to
produce two numbers would be OI-41's defect reintroduced for the sake of
tidiness, so _unparsed_counts became _record_counts and folds both.
_fold_record splits the per-record work out to stay under the complexity
ratchet.

The two numbers in that block are deliberately sourced differently:
unparsed matches note markers (shared constants, so wording cannot drift);
resolution_gaps is recomputed from each record's language_breakdown against
the live tables, because deriving cannot go stale the way string matching can.

Verified end to end

A repo holding Go, Java and Scala:

notes:  go   — limited extraction across 1 file(s) — ...T1... ...T2...
        scala — limited extraction across 1 file(s) — no tree-sitter grammar...
        (java: silent)
manifest: "resolution_gaps": {"go": 1, "scala": 1}

Rescan

DETECTION_VERSION 16 → 17 — notes are record content. 3.1.0 shipped
detection 13, so 14 through 17 all fall in the same unreleased window:
still one rescan.

Gates

1098 tests (6 new), 148 mutants (2 new, both killed non-vacuously), ruff,
mypy --strict, complexity ratchet, fingerprint gate — green locally. Catalogue
budget 148 → 150; both new mutants are ordinary, so the slow tail stays at two.