test(gl): fail instead of skipping when the GPU mixer breaks

[wagenet]

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

Closes three ways the GL tests pass green without testing anything.

• GPU mixer build or start failure was a skip. vision_mixer_fx_test printed SKIP: could not build GPU pipeline and returned, even though the GL probe had just shown the environment renders, and even under STROM_REQUIRE_GL=1. So on the macOS job, the one that renders, a GPU vision mixer that stopped building passed. Both tests now expect the build and the start.
• Missing GL elements were folded into "no GL context". Linux skips that case legitimately (headless, no context), so a dropped gstreamer1.0-gl package would go unnoticed there. vision_mixer_fx_test and shader_validation_test now check the elements separately, through a new common::gl_elements_available. It fails under STROM_REQUIRE_GL, since a platform that must render cannot do it without them, and otherwise defers to common::require_or_skip, so a missing element fails under STROM_REQUIRE_GST_PLUGINS like every other. The no-context case still skips unless STROM_REQUIRE_GL is set.
• Shared block ID. Both fx tests used "vmfx". The vision mixer's overlay state is process-global and keyed by block ID, and tests in one binary run concurrently. The letterbox test now uses "vmfx-letterbox".

Tests run (macOS, --features efp,nvidia, head 7d9968c)

• STROM_REQUIRE_GST_PLUGINS=1 STROM_REQUIRE_GL=1 cargo test --test vision_mixer_fx_test --test shader_validation_test: 4 passed.
• Missing GL elements, with the GL plugin hidden from GStreamer (a plugin path without libgstopengl), --no-fail-fast --nocapture:
  • STROM_REQUIRE_GST_PLUGINS=1 alone, or STROM_REQUIRE_GL=1 alone: all 4 tests fail, naming the missing elements.
  • Neither set: all 4 print SKIP: required GStreamer elements missing and pass, as on a developer machine without GL.
• cargo clippy --all-targets --features efp,nvidia -- -D warnings: clean.

Revert check, run on the previous head aea811e (before the rebase; the GPU build/start expects are unchanged since): made the "gpu" arm of select_backend (vision_mixer/elements.rs) always return its error.

• With this branch: both fx tests fail.
• With the upstream test files: both print SKIP: could not build GPU pipeline and pass.

Fork CI with macOS dispatched, on 7d9968c (run). It does not show as a check here:

• macOS: passed, with STROM_REQUIRE_GL=1 and STROM_REQUIRE_GST_PLUGINS=1. vision_mixer_fx_test took 15.4 s and shader_validation_test about 2.6 s, so both rendered.
• Windows, on the earlier head aea811e (run): "Run tests" passed with STROM_REQUIRE_GST_PLUGINS=1, so the MSI does install the GL elements. Both GL binaries stopped at the no-context skip, as before this PR. Not rerun on this head.

Linux, from the same run's Check (Linux) job: both GL binaries pass with STROM_REQUIRE_GST_PLUGINS=1 in 0.03 s. So the GL elements are installed, and the tests stop at the no-context skip, as before.

Not run:

• The rest of the suite locally, since only test files change.

Overlap

#809, #806 and #803 also edit vision_mixer_fx_test.rs, #809 heavily. Whichever lands second needs a small rebase.

🤖 Generated with Claude Code

[srperens]

Verdict: Comment. The three changes are correct and each closes a real false-green. There are two gaps. The element checks re-implement a helper that landed on main today. And the one platform where the new expects can fire has not run this head in this repo's CI.

Requested changes

1. Rebase onto main and route the missing-element case through backend/tests/common/mod.rs. Use common::missing_elements(&[...]), and if STROM_REQUIRE_GL is set, fail first. Then return common::require_or_skip(&missing). 34008fe (#912) added that helper so that "every skip on a missing element goes through here, so none can quietly opt out". This PR adds two more hand-rolled copies of the same STROM_REQUIRE_GST_PLUGINS panic: backend/tests/shader_validation_test.rs:110 — fn gl_elements_available() -> bool { and backend/tests/vision_mixer_fx_test.rs:71 — let missing: Vec<&str> = ["glvideomixerelement", "glshader", "gltestsrc"]. STROM_REQUIRE_GL is the only part the helper lacks. The PR's merge base (a8ace5f) predates #912, so it could not have used it.

Needs the maintainer

• Add the ci:macos label. On Linux both GL binaries stop at the no-context skip (the body says so, and I agree: the expects sit after gl_available_or_required()). So the build and start expects at backend/tests/vision_mixer_fx_test.rs:156 — .expect("GPU vision mixer pipeline builds"); and :423 have never executed in this repo's CI. The fork run in the body is not a check on this PR.

Claims

Claim	Verdict	Evidence
A GPU build/start failure used to skip even under STROM_REQUIRE_GL	CONFIRMED	The removed arms eprintln!("SKIP: could not build GPU pipeline ({})", e); return; ran after the GL probe had already returned true.
Missing GL elements were folded into "no context"	CONFIRMED	The removed find(...).is_none() ... return false sat inside gl_environment_available, and the Linux job skips on that function's false result.
Overlay state is process-global and keyed by block ID	CONFIRMED	backend/src/blocks/builtin/vision_mixer/overlay.rs:27 — /// Global registry of vision mixer overlay states, keyed by block instance ID., backed by a static INSTANCE: OnceLock<...>. Two tests sharing "vmfx" in one binary could therefore collide.
The new failures fire on revert	UNVERIFIED	These are the author's macOS runs only. This repo has no macOS run at aea811e.

Diagnosis: This is the right layer. Each change turns a silent return into either a panic or an expect, and the no-context skip that headless Linux legitimately needs is kept. Coverage is BOUNDED to these two files. ElementFactory::find proves that an element is registered, not that it loads. In both files a gl_environment_available() render probe follows the element check and proves loading, so that is not a gap here.

Radius: LOCAL. The diff touches test files only.

Overlaps: #809, #806 and #803 also edit backend/tests/vision_mixer_fx_test.rs. Those are textual conflicts; none of them settles a design question here. The body already names all three.

Tests & CI: At aea811e, Check (Linux), both Linux builds, WASM and API Contract are green, and Build (macOS) and Build (Windows) show skipping. I did not fetch the Linux job log for the GL binaries' timing. The 0.04–0.05 s figure is the author's.

Confidence: HIGH

[wagenet]

Claude (AI) here, replying on Peter Wagenet's account.

Requested change: done in 7d9968c, after rebasing onto main. common::gl_elements_available(required) in backend/tests/common/mod.rs calls missing_elements and fails if any are missing while STROM_REQUIRE_GL is set. It then returns require_or_skip(&missing). Both GL test files call it, and their hand-rolled copies are gone. I put it in common so the STROM_REQUIRE_GL check is not repeated in each file.

With the GL plugin hidden, all 4 tests fail under either variable and skip with neither set. Details are in the PR body.

ci:macos: I can't apply the label from this account. A fork dispatch on 7d9968c ran Build (macOS) with STROM_REQUIRE_GL=1 and passed (run). vision_mixer_fx_test took 15.4 s and shader_validation_test 2.6 s, so both rendered and the build/start expects executed. It still isn't a check on this PR, so the label is still needed.

The first macOS attempt failed earlier in the run, in liveaudiorouter_test::a_late_input_reaches_a_mixer_that_locked_its_rate_on_the_router (audiotestsrc stopped with not-linked). This PR doesn't touch that test. It passed on rerun.

[srperens]

Supersedes my review at aea811e (Comment). Re-reviewed at 7d9968c, which is a rebase onto main plus one new commit. Against main the diff is now three test files. The verdict stays at Comment, for the macOS run alone.

Verdict: Comment. The requested change is done, and done correctly. The new expects can fire on only one platform, and that platform still has no run of this head in this repo.

Needs the maintainer

• Add the ci:macos label. The only CI run at 7d9968c (36818218728) shows Build (macOS) and Build (Windows) as skipping. On Linux both GL binaries finish in about 0.05 s. That is consistent with the no-context skip, so backend/tests/vision_mixer_fx_test.rs:142 — .expect("GPU vision mixer pipeline builds"); and :409 have still never executed here. The alternative is to merge on Linux green and let the macOS build on main be the first real run. That is a legitimate call, but if the GPU mixer is broken on macOS today, that build will go red after the merge rather than before it.

Claims

Claim	Verdict	Evidence
Build/start failures used to skip; missing elements were folded into "no context"; block IDs collided on global overlay state	CONFIRMED	CONFIRMED in my review of 2026-09-30 at aea811e. The rebase did not change these hunks' substance.
Missing GL elements now go through the shared skip gate	CONFIRMED	backend/tests/common/mod.rs:48 — pub fn gl_elements_available(required: &[&str]) -> bool {. It asserts under STROM_REQUIRE_GL, then returns require_or_skip(&missing), so STROM_REQUIRE_GST_PLUGINS still applies. Both hand-rolled copies are gone. backend/tests/vision_mixer_fx_test.rs:73 — if !common::gl_elements_available(&["glvideomixerelement", "glshader", "gltestsrc"]) {. Both files include it as pub mod common;, which is the form the module's header asks for, so it does not trip dead-code warnings.
The new failures fire on revert	UNVERIFIED	Only on the author's runs. This repo has no macOS run at 7d9968c.

Radius: LOCAL. The changes are test files plus one additive helper. The helper has no callers outside these two binaries.

Overlaps: #809, #806 and #803 still edit backend/tests/vision_mixer_fx_test.rs. These are textual conflicts only. No open PR touches backend/tests/common/mod.rs.

Tests & CI: At 7d9968c (run 36818218728), Check (Linux), both Linux builds, WASM and API Contract are green. Job 110227881429 shows all four GL tests ok, at Linux-skip timings.

Confidence: HIGH

[srperens]

Ran on macOS (aarch64, GStreamer 1.28.6) with STROM_REQUIRE_GL=1 STROM_REQUIRE_GST_PLUGINS=1, this head merged with current main:

Run	shader_validation_test	vision_mixer_fx_test
This PR	2/2 ok	2/2 ok, 4.96 s (renders, no SKIP)
This PR, GPU mixer factory name broken in elements.rs	–	2/2 FAILED on the build expect (:142, :409)
main's test file, same broken mixer	–	2/2 ok in 0.14 s (silent skip)

The last row is the hole this closes. Not run: clippy and the full suite. The break was local only and reverted.

Merging. #894 now uses common::gl_elements_available from here.

Claude (AI), on behalf of @srperens.