[rmsharp]

SOURCE_EXTS already had .r (R source always counted toward Source LOC), but LANG_MAP had no entry for it, so R never got its own row in the Code by Language breakdown. Found scanning a real R package: 603 .r files, 77,773 LOC — present in Source, invisible per-language.

.qmd (Quarto) and .rmd (R Markdown) were in neither SOURCE_EXTS nor DOC_EXTS, so a file with either extension outside a docs/ path fell all the way through categorize_file's ladder to "other" — not source, not docs, not even LOC-counted (LOC is skipped entirely for "other"). Added both to DOC_EXTS: they're literate-document formats (prose with embedded code), the same bucket .md already lives in.

Also updated the doc-only corpus-disjunction comment (it described a pure-Quarto repo's .qmd as never counted as docs — no longer true; a pure-LaTeX repo's .tex still isn't), and added a new fixture + test that isolates the render-toolchain arm of that disjunction from the doc-corpus arms, since the existing Quarto fixture's own doc_loc now also clears the corpus threshold on its own.

DASHBOARD_VERSION 2.10.2 → 2.10.3 (also independently proposed by #70 and #71 for unrelated changes — whichever of the three merges last will need a fresh bump at that time; not something this PR can pre-empt).

RED-first: all 6 new/changed assertions confirmed failing against unmodified upstream/main before either constant was touched. Full suite 203/203 minus 2 pre-existing failures unrelated to this change (FRAMEWORK_INSTALLED_SOURCE not yet extended for context_budget.py — same baseline reproduced against a clean upstream/main worktree; #71 already addresses it).

---

Edit: addressed all three review findings.

1. Missing CHANGELOG.md entry — added. This was the only one of the five open PRs without one. Confirmed why: the .githooks/pre-commit ledger gate was deliberately bypassed with --no-verify for this upstream-targeted commit (the correct call — that hook enforces this fork's own ledger convention, which doesn't apply to an upstream branch), but the entry itself was simply never written, unlike the sibling PRs. Added now, describing the change, the classification consequence below, and verification.

2. Version collision — now three-way, resolved. #70 and #71 (open the same day) have since moved to 2.10.4 and kept 2.10.3 respectively, so this PR now takes 2.10.5 in both twins + test_dashboard_version.

3. Under-described blast radius — pinned with a new test, not split to a follow-on. You're right that Quarto repos don't actually reclassify (already doc_only via the toolchain-marker fallback; this PR only fixes their reported metrics) but bare .Rmd does — no toolchain marker at all, so adding .rmd to DOC_EXTS is what newly clears the corpus disjunction for that class. Verified by direct reproduction with the same fixture shape (a .R helper + several .Rmd files, no _quarto.yml/_bookdown.yml) against both trees:

upstream/main (pre-fix):  doc_only=False, risks include "No test infrastructure" (HIGH)
this branch (post-fix):   doc_only=True,  that risk replaced by a doc-only advisory

New test test_rmd_analysis_repo_flips_doc_only_and_softens_the_test_risk pins exactly this — RED-verified against a genuine pre-fix scanner checkout (not just git stash, since the underlying fix was already a committed diff on this branch, not working-tree state) before landing.

Re-verified

• python3 -m unittest tools/test_methodology_dashboard.py — 204/204 (198 prior + 6 original + 1 new; the 2 pre-existing failures on this base are fixed by fix(dashboard): extend framework-installed exclusion for context-budget gate #71, not this PR)
• python3 bin/check-links — OK (83 links / 21 files)
• tools/ and starter-kit/ twins confirmed byte-identical
• Rebased onto current main (a2a7275 → post-docs(skills): re-ground the /caveman row's remaining unsupported claim #68-merge) — clean, no conflict (this PR had no CHANGELOG.md entry to collide with before this edit)

🤖 Generated with Claude Code

[KJ5HST]

Ran the checks at the head and verified the premises against main by importing the live constants rather than reading them off the diff. Approving in substance — every claim in this PR holds up, and the tests are the most careful of the five currently open. Two things to fix first, one of them just mechanical.

Premises, all confirmed:

Claim	Verified
.r is in SOURCE_EXTS but absent from LANG_MAP	✓
.qmd/.rmd are in neither set	✓
They fall through categorize_file's ladder to "other"	✓ :287-303
"other" is not LOC-counted at all	✓ :788 — category not in ("assets", "other")
…but only outside a docs/ path	✓ :297 also matches "docs/" in rel_str, exactly as your comment says
Lowercased suffix makes .R/.Rmd resolve to the one key	✓ :775
.tex stays out	✓ absent, so the retained fallback comment is still accurate

Head: unit 201/203 (the 2 are main's, inherited), bin/check-links OK (83 links / 21 files), twins byte-identical.

Specific credit for one thing that's easy to miss. You noticed that the existing Quarto fixture would now clear the corpus arm on its own — silently converting a test that isolated the render-toolchain arm into one that no longer tested it — and added QUARTO_MINIMAL + test_a_pure_quarto_repo_below_doc_thresholds_still_doc_only_via_toolchain to keep the isolation. That's exactly the "incidental coverage evaporates when you improve a fixture" hazard your own #70 is about, caught unprompted in the same subsystem. Nice.

1. The PR under-describes its own blast radius

It's framed as a categorization/display fix. It's also a classification change. Measured across four fixtures, main 2.10.2 vs this head:

=== quarto WITH _quarto.yml ===
  main     doc_only=True   src=0    docs=0     other=6   no_test_risk=none
  #72      doc_only=True   src=0    docs=512   other=0   no_test_risk=none

=== quarto bare (.qmd only, no _quarto.yml) ===
  main     doc_only=True   src=0    docs=0     other=6   no_test_risk=none
  #72      doc_only=True   src=0    docs=512   other=0   no_test_risk=none

=== R Markdown (.Rmd x 5 + a 30-LOC helper.R) ===
  main     doc_only=False  src=30   docs=0     other=5   no_test_risk=HIGH   langs={}
  #72      doc_only=True   src=30   docs=430   other=0   no_test_risk=low    langs={'R': 1}

=== R package with a real test suite ===
  main     doc_only=False  src=360  docs=0     other=1   no_test_risk=medium
  #72      doc_only=False  src=360  docs=82    other=0   no_test_risk=medium

Three things fall out of that:

For Quarto, nothing reclassifies. *.qmd is already a render-toolchain marker (:1735), so a Quarto repo — with or without _quarto.yml — was already doc_only through the fallback arm. What you fix for them is the reported metrics, and that's worth fixing on its own: a Quarto book previously displayed zero documentation and 6 uncounted "other" files. Your in-code comment says this accurately; the title and body read like a misclassification is being repaired.

The reclassification is specifically R Markdown. .Rmd has no toolchain marker — _bookdown.yml is one, but a bare .Rmd analysis project has neither that nor *.qmd. So that class flips doc_only False → True and its "No test infrastructure" risk drops HIGH → low.

I think that flip is correct — an R Markdown analysis project is exactly the research population BL-5/v3.2 exists to score fairly, and your has-tests gate correctly protects the fourth fixture (an R package with tests/ stays code, and now gets its R language row). But a change that moves repos between scoring regimes and suppresses a HIGH advisory is a user-visible verdict change, and nothing pins it: the six new tests cover categorization, language rows, and the preserved toolchain arm — none asserts the downstream classification or risk consequence. Add a test that pins the .Rmd-analysis-repo classification and say it in the body, so a future edit to DOC_EXTS can't silently un-flip a whole population.

2. The version collision is three-way (blocking, mechanical)

You, #70 and #71 all bump DASHBOARD_VERSION 2.10.2 → 2.10.3 and all rewrite the same test_dashboard_version assertions. I merged all three onto origin/main in a scratch worktree:

• all three land, 204/204 green (197 base + 1 from feat(dashboard): document and pin the doc-only detection thresholds #70 + 6 from yours), twins byte-identical
• the only conflict is CHANGELOG.md, and it's between feat(dashboard): document and pin the doc-only detection thresholds #70 and fix(dashboard): extend framework-installed exclusion for context-budget gate #71 — yours carries no ledger entry, so it conflicts with nothing

So the three are code-compatible; two of them just need renumbering to 2.10.4 / 2.10.5 in both twins plus the assertions. I'm planning to take #71 first since it's the branch that un-reds main. (I told #70 "whichever of you lands second needs 2.10.4" before this PR existed — that was one short.)

3. No CHANGELOG.md entry

This is the only one of the five open PRs without one:

#68: CHANGELOG.md  starter-kit/RECOMMENDED_SKILLS.md
#69: CHANGELOG.md  starter-kit/HANDOFFS.md
#70: CHANGELOG.md  starter-kit/…py  tools/…py  tools/test_…py
#71: CHANGELOG.md  starter-kit/…py  tools/…py  tools/test_…py
#72:               starter-kit/…py  tools/…py  tools/test_…py

It changes tracked content, which is what the v3.1 .githooks/pre-commit gate refuses without a co-staged ledger entry — so either the hook isn't enabled on that clone or it got bypassed. FM #27, and the class Phase 0 reconcile-on-read exists to catch. Worth checking git config core.hooksPath .githooks on that clone either way, since a silently-disabled hook will keep doing this.

Completeness sweep — clean

Ran the Learning #10 sweep for sites this could have made stale. No live operative doc describes DOC_EXTS, SOURCE_EXTS, or "Code by Language" — nothing to update. The two places that still say .qmd isn't counted as docs are both frozen dated records (root CHANGELOG.md:568 and CLAUDE.md:118, v3.2 release narration), which the v2.7.1 convention keeps verbatim by design. The one live .qmd mention, RESEARCH_DOCUMENTATION_WORKSTREAM.md:337, is the toolchain matrix's render-command column and is unaffected by categorization.

---

Add the ledger entry and the classification test, take whichever version number falls out of the merge order, and I'll merge this. If you'd rather ship the classification test as a follow-on, say so — it's the one item here I'd accept splitting, since the behavior it pins is behavior I think is right.