test: share the block install harness and bus readers

[wagenet]

Opened by Claude (AI), on Peter Wagenet's account.

Refs #912

Problem

Fourteen tests install one block into a bare pipeline by hand, and about twelve read the bus until a deadline, each in its own way. Every copy has to be fixed separately when the block contract changes.

Change

Moves the copies in seven test files and common/recorder.rs into backend/tests/common:

• common::block: context(), install(pipeline, built) (adds the elements, links internal_links), install_except for the test that leaves the sink out on purpose, and run_setups(ctx), which runs the element setup hooks with one flow ID, as the manager does.
• common::bus: first_error, first_error_until (main's early exit in the SRT program-change test) and collect_errors.
• common::free_udp_port(), replacing three identical srt_port copies.

One commit on main, independent of #935. The only overlap is the pub mod lines in common/mod.rs; whichever merges second needs a trivial resolution there.

Differences from the copies it replaces:

• install links with link_pads, which finds a static or already-requested pad by name and requests one otherwise. Some copies accepted static pads only. For every builder these tests use, the linked pad pairs match the replaced copy's.
• Setup hooks get one flow ID per pipeline, as in production. Some copies passed a fresh ID to each.
• install is still a copy of the manager's linker, now one copy instead of fourteen. Calling the production linker means exporting it from backend/src, a separate PR.

Evidence

Mutations in the test code, run on this head:

• install without its link loop fails mpegtssrt_input_program_change_test, mpegtssrt_streamheader_test, videoenc_profile_negotiation_test, whip_idle_slot_test and whip_slot_caps_reuse_test. The other three binaries need no links.
• run_setups as a no-op fails two tests in output_refusal_test and two in recorder_test.
• collect_errors dropping every error fails the three tests in output_refusal_test.
• first_error and first_error_until dropping every error fails nothing: their one user asserts there is no error. The copy they replace had the same gap.

Tests

Ran (macOS, 2026-10-09): the seven touched test files and recorder_test: 44 passed, 0 skipped. cargo fmt --check and cargo clippy --all-targets -- -D warnings are clean. Not run: Windows, the full suite.

CI: the Linux check failed in vision_mixer_input_format_test::gpu_videotestsrc_direct, a graphics test this PR does not touch; that binary failed in other PRs' runs today. Cargo stopped there, so whip_idle_slot_test and whip_slot_caps_reuse_test did not run in that job (both ran and passed in #935's). The other touched binaries ran and passed. A re-run is pending.

Out of scope

Eight test files still link blocks by hand: drop_on_latency_test, liveaudiorouter_test, efpsrt_data_roundtrip_test, live_recorder_test, whip_video_damage_test, whip_slot_flush_test, whep_fanout_bench, whip_slot_decoder_reuse_test. The first and third also keep their own free-port code. shader_validation_test and vision_mixer_fx_test are left alone too; #917 has merged, so they can move in a follow-up.

🤖 Generated with Claude Code

[srperens]

Noted as a draft, so I am holding off on a full review until it is marked ready for review. If you want one before then, say so here or request me as a reviewer.

[srperens]

Verdict: Comment. The refactor checks out and the touched binaries passed in CI. One check is red, in a test this diff does not touch. Re-run it before merging: gh run rerun 37962999488 --failed. If it comes back green, I see nothing that blocks this.

Claims

Claim	Verdict	Evidence
install links the declared internal links through link_pads, request pads included	CONFIRMED	backend/tests/common/block.rs:51 — src.link_pads(from.pad_name.as_deref(), sink, to.pad_name.as_deref())
The recorder helper now shares the same setup runner	CONFIRMED	backend/tests/common/recorder.rs:66 — crate::common::block::run_setups(&self.ctx);. Every file that includes common/recorder.rs through #[path] (recorder_test, recorder_stall_test, live_recorder_test) also declares pub mod common;, so crate::common::block resolves in all three
The touched binaries ran and passed in CI	CONFIRMED	Check (Linux) job 113930309722: mpegtssrt_input_program_change_test 2 passed, mpegtssrt_streamheader_test 1, output_refusal_test 4, recorder_test 9, rtmp_output_test 16, videoenc_profile_negotiation_test 7. whip_idle_slot_test and whip_slot_caps_reuse_test never ran because cargo stopped at the failure
The red check is not this diff's	CONFIRMED	It is gpu_videotestsrc_direct in vision_mixer_input_format_test, which failed to reach PLAYING. The diff touches no vision mixer file. Check (Linux) passed on main's last five runs, and the same binary has failed on other PRs (#898)

Diagnosis. This moves duplicated code into one place and changes no behaviour, apart from three differences the body names:

• link_pads replaces the copies that used static pads only;
• setup hooks get one flow ID per pipeline;
• install_except covers the test that leaves its sink out on purpose.

The body's mutation results do the guard work a refactor can carry. first_error and first_error_until are unguarded, and the body says so honestly: their only caller asserts that there was no error. install is still a copy, not the production linker. Production feeds internal links through Link strings (backend/src/gst/block_expansion.rs:142 — for (from_ref, to_ref) in build_result.internal_links {), so these tests do not cover how the manager resolves pads. That gap predates this PR, which now keeps it in one place instead of fourteen.

Overlaps.

• #935 conflicts with this PR in backend/tests/common/mod.rs. I checked with git merge-tree: it is the pub mod lines, as the body says. Whichever merges second needs a one-line resolution. #935 already has my Approve, so landing #935 first and rebasing this one is the cheaper order.
• #1067 also edits common/mod.rs and merges cleanly with this PR.

Radius. LOCAL: test code only. Nothing under backend/src changes.

Tests & CI. At dd958b0: both Linux builds, WASM, API Contract and Static JS pass. Check (Linux) fails as above. macOS and Windows were skipped. The code is portable test code, so the main-branch build covering them is acceptable. The branch merges cleanly with current main.

Confidence: HIGH