Test-infra/robustness: renderer visual tests are order-dependent (pass isolated, fail under --gtest_shuffle) — cross-test GPU/renderer state leak (the #530 species) #549

Summary

Several renderer visual-evidence / golden tests pass in isolation but fail under --gtest_shuffle — i.e. they are order-dependent. That is the #530 bug species (renderer/GPU global state that survives a boundary which should reset it), manifested at the test level: some state leaks between tests despite the RendererAttachedTest per-test GLStateGuard(Restore) + renderer reset. This is both a CI-reliability risk (a shuffle, or simply adding/reordering a test, reddens the build) and a strong signal of a real renderer state leak that isn't reset on a scene/path/settings boundary in production.

This is distinct from #530 (that issue is the specific blackboard-fingerprint-cache-survives-ResetTopology instance, whose symptom is a whole-graph cull → empty frame). The failures here are subtler (partial frame corruption, a brightness-ordering flip, a shadow-golden mismatch), so they are other instances of the same species — though triage should check whether #530's fix incidentally resolves any of them.

Evidence (fresh Debug build, 2026-07-04)

Full suite --gtest_shuffle --gtest_repeat=2: iteration 1 → 2 failures, iteration 2 → 4 failures. Every shuffle-failing test passes when run alone (verified, each in its own process):

(AppLaunchSmoke.OloRuntimeLaunchesCleanly also failed — that one is a false failure from a partial build that didn't produce OloRuntime.exe, ignore.)

Deterministic reproduction

The pollution reproduces within just the renderer visual suite (246 tests), so the leaking state is inside that set:

build/OloEngine/tests/Debug/OloEngine-Tests.exe \
  --gtest_filter="*VisualEvidence*:*Golden*:*Occlu*:*ProceduralSky*:*Evidence*:*GTAO*:*Water*:*SSAO*:*SSGI*:*SSR*" \
  --gtest_shuffle --gtest_random_seed=58251
# => OccludedInstanceFieldScene.OcclusionDoesNotCorruptOrHoleTheFrame FAILS
# seed 58250 => GoldenImageTest.SceneShadowIntegrationGolden FAILS; 58252 => Occlusion FAILS again

Each of those same tests passes with --gtest_filter=<TestName> alone.

Notes / narrowing

• The victim tests span occlusion culling, shadow golden, procedural-sky bake, and GTAO/EASU → a shared renderer/GPU global that isn't reset between tests. Candidate leaked state: Renderer3D::s_Data fields not zeroed by the per-test reset (Settings.Path / ActiveGraphPath, PostProcess.ActiveAOTechnique, upscale mode, OcclusionResultsAvailable, OcclusionHZBValid), the occlusion singletons (OcclusionQueryPool / OcclusionStateManager / persistent HZB pyramid), IBLCache, or raw GL enable state.
• GL clue: the run logs 4× GL_INVALID_ENUM error generated. <cap> enum is invalid (a glEnable/glDisable with an out-of-range capability) — a GL-state-handling smell worth checking as part of the same audit.
• The immediate predecessor of the occlusion failure (seed 58251) was a 0 ms structural test, so the leak is cumulative, not a single adjacent polluter — a proper bisection (shrink the filter around the fixed seed) is needed to name the exact producer.

Fix direction

1. Bisect with the fixed seed (58251) to identify the producing test, then find which global it leaves dirty.
2. Extend the RendererAttachedTest teardown to reset that state (and audit Renderer3D::s_Data for every "valid until invalidated" member against every reset path — same audit that surfaced RenderGraph: second consecutive Deferred-path switch in one process culls the entire graph (reads=0/writes=0) #530), and/or reset the occlusion singletons between tests.
3. Add a shuffled renderer-visual job to CI so this regresses loudly rather than depending on definition order.

Score

capability: 3
craft: 1
stability: 5
decay: 3
effort: 3
confidence: 0.8
learning: 3
fun: 3
kano: table-stakes
blocked_by: []
blocks: []

_{Rated per issue-scoring · score = confidence × (capability + craft + stability + decay) / effort, derived by the picker.}