Make npm run cov independent of node --test default discovery - #1512

Approving. This fixes a real, reproducible defect, and I can confirm the fix works on the exact Node version where the old command was vacuous.

The headline result

My last several reviews carried a standing caveat: npm run cov reports a meaningless 100.00 over 0 tests in my environment, so coverage is unverifiable. That caveat is now closed. Node v23.11.0, same machine, same node_modules:

So naming fjs/emergent_testing/all.test.ts as the entrypoint does exactly what the TODO says it does, including on the one released Node line where default discovery finds nothing. I will stop excluding coverage from verification.

Verification battery (at 65c886a7, clean tree)

• npx tsc --noEmit — exit 0.
• npm run prepack from a freshly cleaned tree — exit 0.
• npm test — pass: 2495, fail: 0, total: 2495, unchanged from main.
• npm run ci-update then git add -A && git diff --cached --exit-code — clean, so the generated .github/workflows/ci.yml really is the output of the edited coverageInclude const. deno.json is not generated (it is only referenced in a comment), so its hand-edit is the right way to keep the third copy in sync.

Claims re-derived rather than assumed

• Dropping **/module.f.ts and .f.ts from the Deno filter is safe. git ls-files | grep -cE '\.f\.ts$' is 0 at this head — no authored .f.ts remains, so neither filter can lose a module.
• The unquoted glob in package.json is not shell-expanded. --test-coverage-include=**/module.f.mjs is one word; sh would need a directory literally named --test-coverage-include=* to match, so it passes through intact. (There is a depth-1 match, fjs/module.f.mjs, so this was worth checking.)
• The 2431-vs-2495 explanation in the TODO reproduces here. Of the 2431 top-level ✔ lines on v23.11.0: 2376 carry the ... inline marker and the remaining 55 are exactly the .throw tests, with no others in that set. That is the table the doc states, measured independently on a third Node version.
• The rewritten star comment in fjs/emergent_testing/module.f.mjs is accurate. usesInlineTestContext returns true for bun unconditionally and for node below the 26.0.0 baseline, which is what the new comment says; the old comment also described the marker as ' *' when the code has emitted ' ...' (const star = o.inlineTestContext ? ' ...' : ''), so this corrects a stale comment as well.

CHANGELOG entry is present, sits in Unreleased, and links only /pull/1512 per §8.3. The TODO rewrite is honest about which part of the earlier diagnosis was right (discovery) and which framing was wrong («vacuous for several PRs» — it was vacuous only off-CI, on v23), and keeps both rather than quietly deleting the wrong one. I like that.

Nit, non-blocking: the CHANGELOG sentence «The dead .f.ts extension leaves the Node and Deno coverage filters» parses as departs from on the second read but as leaves behind on the first. «is dropped from» would be unambiguous.

Nothing blocking.

Addendum — re-verified at the new head. My approval was posted at 65c886a7; the head has since moved to 9a15a993, which is not just a merge of main: it adds fddc22ac (CHANGELOG reword) and 959b7650 (run deno task cov in CI instead of inlining the command). I re-ran the battery at 9a15a993 and it still stands — re-approving in effect, nothing new to raise.

959b7650 is a genuine improvement over what I approved. Removing coverageInclude leaves deno.json as the single owner of the Deno filter, which is the same shape the Node side already has (package.json owns npm run cov), so the "keep these two semantically equal" comment I read as a maintenance hazard is now a one-owner invariant instead. The public-surface removal is correctly marked **BREAKING CHANGES:** per §8.4 and links only /pull/1512. The CHANGELOG reword also fixes the "leaves the … filters" ambiguity I nitted — thanks.

At 9a15a993:

• npm run ci-update then git add -A && git diff --cached --exit-code — clean, so the regenerated ci.yml step really is deno task cov.
• grep -rn coverageInclude fjs/ outside todo/ — no hits, so the export is gone with no dangling importer (fjs/ci/proof.f.mjs dropped its import too).
• npm test — pass: 2495, fail: 0, total: 2495 against 2496 on current main (20cd6aa9). The −1 is exactly the retired coverageInclude proof case; the merge of #1509 contributes +1, so 2495 + 1 − 1 = 2495. Accounted for, not a regression.
• npm run cov on Node v23.11.0 — tests 2431 / pass 2431 / fail 0, all files 99.93. The headline result is unchanged: still a real coverage report where main reports tests 0.