test(gpu): focus evidence lanes on behavioral coverage

[morluto]

Description

The GPU evidence integrations accumulated tests that mirrored private helpers, repeated stronger end-to-end coverage, or required source edits whenever internal structure moved. That made the adapter lane slower and noisier without proportionally strengthening the contracts introduced around #131–#136.

This change refocuses those lanes on observable behavior: immutable artifact preservation, bounded imports, failed-attempt recovery, extraction outputs, public capture and capability services, MCP transport behavior, and provider compatibility boundaries. It removes 23 collected cases while retaining or strengthening the failure modes and compositions those issues depend on.

Approach

• Move managed compiler and sanitizer capture compositions to application-owned process tests, while keeping parsing and collection behavior with the adapters.
• Replace duplicated and private-structure checks with canonical validation documents, fake official Nsight interfaces, process simulations, and public service results.
• Strengthen provider readiness so Compute Sanitizer requires nvcc and Nsight Compute requires the official ncu_report interface before a live lane can run.
• Exercise NVBench partial-output modes through the managed capture boundary and preserve the real sidecar/extraction proof in the optional provider lane.
• Remove the unused kernel-build bundle_paths test seam and keep provider-specific bundle bounds with the provider and importing service rather than the generic request model.
• Accept the full NVIDIA Compute Sanitizer version banner emitted by current installations.

Suggested review order: start with the application-owned capture tests and ownership changes, then review the adapter fixture reductions, and finish with the four small production cleanups.

Commands run

uv run ruff check src tests
uv run ruff format --check src tests
uv run mypy src tests
uv run lint-imports
uv lock --check
uv run python tools/test.py ownership
uv run python tools/test.py collection
uv run python tools/test.py adapters
uv run python tools/test.py process
uv run pytest -q
uv run python tools/test.py optional-compute-sanitizer
FLAMEOX_NVBENCH_EXECUTABLE=$NVBENCH_EXECUTABLE uv run python tools/test.py optional-nvbench
uv run pytest tests/adapters/test_rocprofv3_integration.py -q

Results:

• Full deterministic suite: 854 passed, 192 deselected.
• Collection and ownership receipts: 1,046 node IDs across 99 files; 109 ownership records.
• Adapter lane: 262 passed, 20 deselected in 20.806s, down from 28.253s on main (26.4% faster).
• Process lane: 143 passed, 1 deselected in 155.097s, down from 164.009s on main (5.4% faster) while adding nine application-level compositions.
• Live Compute Sanitizer lane: 2 passed on an RTX 3060.
• Live NVBench/Nsight lane: 2 passed using NVIDIA/nvbench at c18488992e313240166f588b9ee4da3e0de76004, including binary sidecar extraction and a successful Nsight counter capture.
• ROCm-shaped Perfetto integration: 1 passed.

Compatibility

There are no public API, protocol, schema, or storage-format changes. The generic bundle request no longer carries a provider-specific Pydantic member limit; the importing service still enforces the hard 100-member boundary, and typed providers retain their stricter validation.

Compute Sanitizer, NVBench, and Nsight Compute received live NVIDIA evidence on this host. Triton and CuTe remain fixture/process-simulation backed in this change. The ROCm test uses a project-owned ROCm-shaped Perfetto fixture and does not claim native rocprofv3 PFTrace compatibility or AMD hardware proof.

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

[morluto]

@codex

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Chef's kiss.

Reviewed commit: d021ec0f3e

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".