What

Summary

Round 3 follow-up to PR #1108 (Round 2 of issue #1106). Two changes:

1. Guard 2 coverage-flaky tests in tests/unit/http2_client_branch_test.cpp (SlowWriteServerSettingsStillCompletesHandshake, EchoOneTruncateAtNineDropsResponsePayloadFailingGet) with #ifndef NETWORK_COVERAGE_BUILD. tests/CMakeLists.txt defines that macro on the branch-test target only when ENABLE_COVERAGE=ON, so Debug/Release matrix runs continue to exercise the slow_write partial-read accumulator and truncate-on-connected request-error branches, while the coverage workflow no longer surfaces them as measured-failing tests.

2. Add 7 dispatcher TEST_F cases that drive previously-unreachable handlers in process_frame by writing server-originated frames after the SETTINGS handshake. Uses a new post_handshake_frames hook on mock_h2_server_peer (Phase 2E.R3 substrate, first commit) that bypasses the injector so production frame_class::serialize() output reaches the client unmodified.

Change Type

• Test only (no src/ changes)
• Feature
• Bugfix

Affected Components

Why

After PR #1108 merged, the post-merge coverage run 25449908024 at b4692fed showed http2_client.cpp line/branch coverage unchanged at 18.8% / 9.9% (target 60% / 40%). Issue #1106 was therefore left OPEN with a recommendation comment proposing exactly the two changes shipped here:

1. Stabilise the two positive-path TEST_F under coverage instrumentation (option B: NETWORK_COVERAGE_BUILD ifdef).
2. Add a fault-injection test specifically targeting process_frame dispatch — currently no test in http2_client_branch_test.cpp exercises the connected-state frame-by-frame branch space (handle_window_update, handle_rst_stream, handle_ping, handle_goaway).

Why these specific dispatcher tests

The 4 negative-assertion tests added in PR #1108 (drop / truncate / malform-type / malform-streamid) all assert EXPECT_FALSE(is_connected()) and exercise only early-connect lines that pre-existing happy-path tests already hit. Net coverage contribution: zero. The 2 positive-assertion tests fail under coverage instrumentation, also contributing zero.

The TEST_F batch added here advances the client past the connected state via the unmodified happy-path SETTINGS handshake (which works under coverage instrumentation) and then has the peer write one precisely-formed server-originated frame. Each frame routes through process_frame to a specific handler:

Where

• Test-only scope. src/ is not touched. No public-API change. Production frame classes (ping_frame, goaway_frame, window_update_frame, rst_stream_frame) are reused from internal/protocols/http2/frame.h so wire format always matches what frame::parse expects.
• The new TEST_F batch sits below the Phase 2E.R2 group under a dedicated section header marked Phase 2E.R3 (Issue #1106).

How

Substrate change (commit 1)

mock_h2_server_peer gains a fourth constructor argument:

explicit mock_h2_server_peer(
    asio::io_context& io,
    reply_mode mode = reply_mode::drain_only,
    injection_spec inject = {},
    std::vector<std::vector<std::uint8_t>> post_handshake_frames = {});

Each entry is written verbatim immediately after settings_exchanged_ flips true and before the per-mode branch. The injector is bypassed (this hook is for tests that need exact wire format, not fault injection). All existing call sites compile unchanged because the new parameter defaults to empty.

Test additions (commit 2)

Each new TEST_F follows the same pattern: serialize a single server-originated frame via the production class, construct the peer with that frame in post_handshake_frames, complete the SETTINGS handshake via make_connected_client, wait briefly for run_io to consume the frame, and assert the resulting state.

Local verification

Test plan for reviewers

cmake --build build --target network_http2_client_branch_test -j 4
./build/bin/network_http2_client_branch_test \
  --gtest_filter="Http2ClientHermeticTransportTest.Server*:Http2ClientHermeticTransportTest.SlowWrite*:Http2ClientHermeticTransportTest.EchoOneTruncate*"
# CI expectation: 9/9 PASS under Debug/Release; 7/7 PASS under coverage
# (the 2 ifdef'd tests are excluded from coverage builds by design)

Acceptance follow-up (#1106)

Coverage targets (line >= 60%, branch >= 40% on http2_client.cpp) are measured by the post-merge coverage workflow on develop. This PR's CI verifies test correctness; the coverage delta will appear on the next develop coverage run after merge. Whether one PR can close the gap entirely or whether a Round 4 is needed will be determined from that measurement, following the Round 1 / Round 2 convention of keeping #1106 open until coverage is verified.

Breaking changes

None. The mock_h2_server_peer constructor change is additive with a default value.

Rollback plan

Revert this PR. Production code is not touched. The substrate change is independent of any caller, so no other tests need adjusting.

Related

• Closes test(http2): expand http2_client.cpp error-branch coverage using Phase 2 substrate (round 2) #1106
• Part of Expand unit test coverage from 40% to 80% #953 (Round 3 of test-coverage epic)
• Builds on PR test(http2): add frame_injector-driven error coverage round 2 (#1106) #1108 (Round 2, merged 2026-05-06)
• Substrate provided by: PR test(support): add frame_injector for protocol-aware fault injection (Phase 2E of #1074) #1105 (Phase 2E of infra(test): protocol-aware loopback peers + error injection (Phase 2 of #1060) #1074, merged 2026-05-06)

Checklist

• Code follows existing tests/unit/http2_client_branch_test.cpp style
• Self-review completed
• New tests added (7 TEST_F)
• No src/ changes
• No sensitive data exposed
• Commits are atomic and well-described
• Related issue linked (Closes test(http2): expand http2_client.cpp error-branch coverage using Phase 2 substrate (round 2) #1106, Part of Expand unit test coverage from 40% to 80% #953)