Collect server subprocess coverage and report production-focused coverage gaps #534

Tracking plan: #528
Priority: P1 for coverage integrity. Evidence: exact-audit-revision LCOV artifacts, with profile-loss cause to verify.

Problem and evidence

The coverage baseline at 4e523ba reports 80.81% Linux / 78.23% Windows line coverage and 84.81% / 81.57% function coverage. Nevertheless, both artifacts record zero execution for the transport loop, reply writers, and RPC entry handlers, despite black-box server tests exercising some of those paths.

The test client kills the server on Drop, which likely prevents normal LLVM profile flushing. Verify that cause rather than equating zero recorded coverage with absence of tests.

Aggregate coverage also includes inline test code. Restricting JSONRPC orchestration LCOV to executable lines before its test module gives 415/900 covered lines (46.1%) on both measured platforms. Conda/Poetry subprocess and diagnostic paths contain substantial unrecorded coverage. There is no macOS coverage job, although macOS functional tests exist.

Scope

• Collect instrumented server-subprocess profiles via normal bounded shutdown, unique profile output names, and inclusion in the final report.
• Add production-focused and changed-code coverage reporting; retain existing whole-workspace metrics for continuity rather than silently changing their meaning.
• Add focused failure-path/branch tests for transport and handler behavior and measure macOS-specific runtime paths with a macOS coverage job.
• Keep exact-base comparisons and fail-closed behavior. Do not optimize for a headline percentage by adding trivial tests or excluding difficult production paths.

Acceptance criteria

• A known black-box request visibly increases coverage in the corresponding handler, transport loop, and response writer on supported platforms.
• A normal server shutdown flushes/collects the expected profile; missing subprocess profiles are diagnosed rather than treated as success.
• Reports distinguish production code from inline tests and expose uncovered changed production paths.
• Targeted branch/error-path reporting is introduced where supported, with documented tool/platform limits.
• macOS coverage is measured, or a concrete toolchain blocker is recorded and split into a linked follow-up before this item is closed.
• Comparator changes and any baseline/schema transition have tests; exact-base line/function regression protection is preserved.

Dependencies and prior work

Depends on #529 for clean shutdown. Coordinate transport fixtures with #532. #522 / #524 is the existing fork-quality-workflow prerequisite. This extends the completed coverage expansion in #389 and gate enforcement in #503; it does not reopen those completed tasks or assert that the current gate is absent.