ci(coverage): disable jacoco PR comments, relax overall delta threshold to 0.4%

[bladehan1]

What

Three targeted fixes to the coverage-gate CI job in pr-build.yml:

1. Add comment-type: summary to both jacoco-report steps — routes output to the workflow step summary instead of PR comments.
2. Relax MAX_DROP from -0.1% to -0.4%.

---

Why

Issue 1 — 403 errors on PR comment posting

The coverage-gate job declares only permissions: contents: read. The madrapps/jacoco-report action defaults to comment-type: pr_comment, which calls the GitHub PR comment API and requires pull-requests: write. Every CI run was producing a 403 error in the logs, though it did not block the gate because the action's own continue-on-error default is true.

Setting comment-type: summary makes the action write to $GITHUB_STEP_SUMMARY (a local file write, no API call), so the 403 disappears entirely without needing to grant broader token permissions.

Issue 2 — coverage-changed-files reports INSTRUCTION coverage, not LINE coverage

The value shown as Changed Files Coverage is bytecode instruction coverage, not source-line coverage. This is a fixed behavior of madrapps/jacoco-report@v1.7.2 — the action has no coverage-type parameter and internally reads only the INSTRUCTION counter from JaCoCo XML.

Why INSTRUCTION ≠ LINE and why it matters:

• JaCoCo XML records six counter types per source file: INSTRUCTION, BRANCH, LINE, COMPLEXITY, METHOD, CLASS.
• A single source line compiles to multiple bytecode instructions. For example, a null-check if (x != null) may emit 3–5 instructions; a fluent chain on one line may emit 10+.
• INSTRUCTION coverage is therefore sensitive to compiler expansion: generated code (constructors, getters/setters, toString, enum boilerplate) inflates the instruction count without adding meaningful semantic coverage.
• When a PR touches one method in a class that also has many generated accessors, the changed-file INSTRUCTION coverage can read significantly lower than the LINE coverage a reviewer would intuitively expect.
• Concretely: the 93.3% figure visible in recent CI runs is INSTRUCTION coverage of the diff files, not line coverage.

To switch to LINE coverage the action would need to be replaced with a custom script that reads the LINE counter (<counter type="LINE" covered="N" missed="M"/>) from each matching <sourcefile> node. The action itself does not expose this as a configuration option in any released version through v1.7.2.

Issue 3 — -0.1% delta threshold causes false failures

Normal variance in test execution timing, minor refactoring (renaming, extracting constants), or adding tests to previously uncovered edge-case paths can shift overall INSTRUCTION coverage by ±0.2%. A threshold of -0.1% triggered spurious gate failures in these scenarios. -0.4% still catches real coverage regressions while eliminating the noise.