What

After three rounds of expanding tests/unit/http2_client_branch_test.cpp with substrate-driven TEST_F (PR #1108 added 6, PR #1109 added 7 — 13 total), the per-file coverage on src/protocols/http2/http2_client.cpp is still 18.75% line / 9.91% branch (108/576 LH/LF, 97/979 BRH/BRF — byte-identical between b4692fe and d537864). Both Round 2 and Round 3 produced exactly 0pp delta on the production source file. Diagnose this before any further TEST_F work.

Why

• Round 1/2/3 acceptance criteria for test(http2): expand http2_client.cpp error-branch coverage using Phase 2 substrate (round 2) #1106 require >= 60% line / >= 40% branch on http2_client.cpp. Current shortfall is -41.25pp / -30.09pp despite 13 new TEST_F.
• The 0pp pattern strongly suggests a measurement or build-system issue, not a test-quality issue. Adding more TEST_F without root-causing this would repeat Round 3's symptom.
• Parent epic Expand unit test coverage from 40% to 80% #953 cannot meet its acceptance criteria while http2_client.cpp (the single largest contributor to the gap) is stuck at 18.75%.

Where

How

Acceptance criteria

• Confirm whether the 7 Round-3 dispatcher TEST_F (ServerPingFrameDrivesHandlePingAndKeepsConnectionAlive, ServerPingAckFrameIsAbsorbedSilently, ServerGoawayFrameFlipsConnectionStateToDisconnected, ServerWindowUpdateOnConnectionStreamExpandsWindow, ServerWindowUpdateOnUnknownStreamIsSilentlyIgnored, ServerRstStreamOnUnknownStreamIsSilentlyIgnored, ServerUnknownFrameTypeIsHandledWithoutCrashing) actually executed under coverage instrumentation in run [25464342500] (grep the Run tests step stdout for each TEST_F name).
• If they did NOT execute, identify which CMake/macro mechanism excluded them (likely NETWORK_COVERAGE_BUILD over-scoping) and submit a fix PR scoped only to the two flaky TEST_F. Re-run coverage workflow on the fix and re-measure.
• If they DID execute but produced 0pp, run gcov locally against http2_client.cpp.gcda from a baseline-only run and a baseline+new-TEST_F run, diff the resulting .gcov files line-by-line, and document which lines are net-newly-hit (or confirm the empty set).
• Verify the lcov filter pipeline does not silently drop a second production-source SF for http2_client.cpp (raw coverage.info had 2 SFs; the second is the test file, but confirm no other --remove rule strips a second production compile unit).
• Output: a one-page diagnostic note in docs/coverage/http2_client_round-4-diagnosis.md (or a comment on this issue) capturing which hypothesis is correct and the proposed fix.

Out of scope

• Adding more TEST_F (a separate Round 5 sub-issue will be opened only after diagnosis is complete).
• Changes to src/protocols/http2/http2_client.cpp itself.
• Cross-file coverage work for other epic-Expand unit test coverage from 40% to 80% #953 contributors.

Hypotheses to investigate (priority order)

1. CMake NETWORK_COVERAGE_BUILD over-scoping — highest probability, easiest to verify, biggest blast radius. If the macro is set on the test target rather than around the two specific TEST_F, the 7 new dispatcher TEST_F never reach the coverage binary.
2. Hit-line overlap with baseline ConnectCompletesSettingsExchangeWithMockPeer — medium probability. If true, the new tests are valuable for behavior-correctness but coverage-redundant; a different mock_h2_server_peer mode would be needed to drive truly-uncovered branches.
3. lcov filter dropping a second production compile unit — low probability after the SF-count check, but a final sanity pass on --remove/--extract rules is cheap.

Test plan for reviewers

After the fix PR (assuming hypothesis 1 lands):

gh run list --repo kcenon/network_system --branch develop --workflow "Code Coverage" --limit 1
# Expect: post-fix run shows http2_client.cpp line/branch coverage moved from 18.75/9.91% upward

Related

• Part of Expand unit test coverage from 40% to 80% #953 (parent coverage epic)
• Part of test(http2): expand http2_client.cpp error-branch coverage using Phase 2 substrate (round 2) #1106 (Round 1/2/3 work; this issue is the diagnostic gate before Round 5)
• Substrate provided by infra(test): protocol-aware loopback peers + error injection (Phase 2 of #1060) #1074 (Phase 2E frame_injector + mock_h2_server_peer)
• Round 1 (closed without target met): test: expand coverage for src/protocols/http2/http2_client.cpp #991, test(http2): expand http2_client.cpp coverage to 80% line / 70% branch #1062
• Round 2 PR: test(http2): add frame_injector-driven error coverage round 2 (#1106) #1108 (merged b4692fe)
• Round 3 PR: test(http2): stabilize coverage and expand dispatcher branches (#1106) #1109 (merged d537864)
• Reference coverage run: https://github.com/kcenon/network_system/actions/runs/25464342500