fix(ci): gate changed-line coverage with diff-cover

[bladehan1]

Summary

Replace the coverage-changed-files gate (whole-file INSTRUCTION coverage) with a true changed-line LINE coverage gate powered by diff-cover, and make the gate robust against non-Java PRs and tool failures.

Commit included

Subject	Purpose
ci(coverage): gate changed-line coverage	Introduce diff-cover as the changed-code gate; switch madrapps/jacoco-report to summary-only; handle non-Java PRs via an NA sentinel reported as SKIPPED; guard empty SRC_ROOTS; verify diff-cover.json was produced; clarify LINE vs INSTRUCTION semantics in the metrics panel.

Why

1. Patch coverage semantics. The previous gate used madrapps/jacoco-report's coverage-changed-files output, which is the whole-file INSTRUCTION coverage of each changed file — not the coverage of the lines the PR actually changed. diff-cover provides the industry-standard "changed-line" measurement.
2. Drop the 403 noise. jacoco-report's default PR comment needs pull-requests: write, which fork PRs do not get. Job logs kept filling up with 403 errors.
3. Stop wrongly failing non-Java PRs. With just diff-cover, a PR that only touches workflows / docs / proto / test code produces total_num_lines=0, and the original jq call exited with 1. The NA sentinel now maps that state to SKIPPED in the enforce step.

Design notes for reviewers

• CI wall-clock impact is ~0. The new coverage-base job runs independently of docker-build-debian11 (they share no needs:). Since docker-build-debian11 is already the slowest build (~16 min with testWithRocksDb), adding a parallel coverage-base of similar length does not extend the critical path. The gate job waits for both, so total PR wall-clock is unchanged.
• Metric mix is intentional and footnoted. Changed-line Coverage uses LINE coverage (diff-cover), while PR Overall Coverage / Base Overall Coverage / Delta use INSTRUCTION coverage. madrapps/jacoco-report v1.7.2 is the latest release and hard-codes the coverage-overall output to INSTRUCTION; there is no coverage-type input yet — see the open madrapps/jacoco-report#82 on milestone 1.8. The METRICS panel and step summary now print an explicit note so readers do not accidentally compare the two counters.
• SKIPPED scope. diff-cover --src-roots is limited to */src/main/java. PRs touching only src/test/java, build/generated (proto output), docs, or CI will be reported as SKIPPED. In practice proto-only PRs with non-trivial generated-code coverage shifts are rare in this repository; we accept the gap.
• No unit tests for the CI script itself. The bash in pr-build.yml is exercised end-to-end on every PR via its own CI run rather than through a shell unit-test framework. The NA / SKIPPED path and the happy PASS path have both been verified end-to-end on prior fork PRs before this submission. Adding bats or similar is possible as a follow-up if desired, but is out of scope here.
• Gate thresholds preserved. MIN_CHANGED=60 (for Changed-line LINE) and MAX_DROP=-0.1% (for Overall INSTRUCTION delta) are unchanged from the existing workflow.
• Failure modes. The new exit 1 branches only fire in genuine error states: SRC_ROOTS is empty (unexpected workspace shape) or diff-cover.json is missing (tool crash). Neither can be triggered by ordinary PR content.

Follow-ups (not in this PR)

• Migrate Overall to LINE once madrapps/jacoco-report 1.8 ships coverage-type, or by self-parsing the JaCoCo XML. Unifies the metrics panel on one counter.
• Add proper shell unit tests (e.g. via bats) for the enforce step, so the three-state logic (NA / numeric / error) is covered without needing a real CI run.
• Revisit comment-type: summary if reviewers prefer on-PR comment visibility; restoring comments requires granting pull-requests: write on the coverage-gate job.

Test plan

• Workflow-only / non-Java PR → Changed-line Coverage: NA, Changed-line Gate: SKIPPED, gate passes. Verified end-to-end on a prior fork PR carrying the same commits.
• Java PR with changed lines → Changed-line Coverage: X%, Changed-line Gate: PASS at X > 60. Verified end-to-end on a prior fork PR (25 changed lines @ 100%).
• Java PR with changed lines at X <= 60 → FAIL. Logic reviewed; path not force-exercised.