bench records: the noise verdict rides the row, and the das backend cell names the armed vulkan tier

[borisbat]

The records batch from the #3687 review discussion — pushed to that branch moments after the merge snapshot, so it rides its own PR (cherry-pick of d51d24b, unchanged).

Both halves are Boris-ruled:

• Records carry the mint's noise verdict. BenchRun gains @optional noise — tune_manifest_noise() reads provenance.noise out of the tune sidecar ("ok" / "overridden", "" when absent), stamped at both das record sites in lcpp_bench (model rows + the ASR cell). A mint taken through a failing noise gate (DAS_TUNE_NOISE_OVERRIDE=1) is now visible on the board, not only inside the sidecar. Documented in METHODOLOGY.md "The record"; round-tripped in test_bench_records_schema.das.
• The das row's backend names the armed vulkan tier. The schema already documented backend : "cpu" | "metal" | "vulkan", but the cell was computed from --ngl alone — a vulkan-armed run published rows labeled cpu. The das row now reports metal (--ngl protocol) / vulkan (moe_gpu_tier_installed()) / cpu; the llama.cpp reference row keeps the --ngl-derived value (llama-bench knows nothing of the vulkan tier). The board oracle already loud-skips backend/flavor pairs no rig leg produces, so vulkan rows are skipped with a warning until a vulkan leg joins the rig.

Validation: schema test 18/18 under -jit, lint 0, formatted, compile-checked.

🤖 Generated with Claude Code

https://claude.ai/code/session_01H3CQmhA327qR2RyKZCKvj6

[Copilot]

Pull request overview

This PR updates dasLLAMA benchmark records to (1) surface the tune sidecar’s noise-gate verdict directly in BenchRun, and (2) correctly label the das engine’s backend as vulkan when the Vulkan tier is armed (instead of deriving it solely from --ngl / Metal).

Changes:

• Extend BenchRun with an optional noise field and populate it from tune sidecar provenance.
• Fix das benchmark row/backend labeling to distinguish metal (--ngl) vs vulkan (armed tier) vs cpu.
• Update methodology docs and schema round-trip tests to cover the new field.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
modules/dasLLAMA/tests/test_bench_records_schema.das	Adds noise to test fixtures and validates JSON round-trip for present/absent noise values.
modules/dasLLAMA/performance/profile_common.das	Introduces tune_manifest_noise() and extends BenchRun schema with @optional noise.
modules/dasLLAMA/METHODOLOGY.md	Documents the new noise record field and its intent on the benchmark board.
modules/dasLLAMA/benchmarks/lcpp_bench.das	Stamps noise into bench records and fixes backend labeling for das rows (including Vulkan tier).

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated no new comments.

Suppressed comments (1)

modules/dasLLAMA/benchmarks/lcpp_bench.das:875

• das_backend is derived from moe_gpu_tier_installed(), which only indicates the Vulkan tier is armed. The tier can be armed-but-idle for a given model (e.g. non-q8 serving mode), in which case the run actually executes on CPU but rows will still be labeled vulkan (markdown + JSON record). Prefer keying the backend off gpu_tier_status().supported (or equivalent “model rides tier” signal) instead of tier installation alone.

        let backend = gpu ? "metal" : "cpu"
        let das_backend = gpu ? "metal" : (moe_gpu_tier_installed() ? "vulkan" : "cpu")
        let gguf = base_name(mpath)