the coverage followup - the bind-number gate, the orchestrator's instruments, the fixture rail

[borisbat]

The kitchen-sink followup ruled after #3909: three items, one PR.

1. The race-arm bind-number gate

check_race_bind_numbers in modules/dasLLAMA/REVIEW.das: every hand-written kn_buffer/kn_kargs literal in a race or knockout arm must land on a binding number its target kernel class declares. The class resolves from the arm's kn_tgmem constant - with or without the kernel's own _msl suffix (the resolver bug the round caught: annotation names carry it, so every race BASE arm was initially invisible; the fixed gate covers 64 of the 94 hand-bind literals) - or from kn_pipeline's pso built by pipeline_from_source in the same function. The licensed residue - arms whose pso/tgmem arrive as function parameters - is stated in the charter and in REVIEW_GPU.md's bounded machine-checked claim. Controls: re-introducing the historic mx4 mis-bind reds, a base-arm poison reds, a kargs poison reds; the clean tree is green interp+JIT. This is the gate for the mis-bind class that silently un-crowned the mx4 tensor twin and halved a board cell - the race's own envelope check fails closed, so a mis-numbered arm just loses and nobody learns.

2. The measurement orchestrator's instruments

The tdd audit of #3909 found the orchestrator side bare; the pure seams now exist and are pinned, each with a must-red control:

• asr_pybench_args per-tool arms (onnx carries the --out recovery file and never a device; nemo forwards --device; clip fan-out, threads/reps literals)
• asr_gpu_pair_tool - the GPU-twin roster extracted from the inline clause, [arch]-cited to #asr-gpu-pairs (per-citer audit: no drift)
• records_run_verdict - the ZERO-cells refusal extracted pure, with the orchestrator's failure arm routed through it so both arms are production-reachable
• canary_qwen_bench.py gains torch-free pure seams (rep_is_degenerate, per-rep build_prompts) behind --self-test, exits 1 on a zero-BENCH-row run, and refuses argless; one live spawn proves a BENCH row (jfk 2558.4 ms, rc 0) with the timed region byte-unchanged
• the crowned word-level whisper twin (followup 60's ruled resolution: per-lane gating stands, the crowned lanes answer to WORD equality) - crowns pinned via the latch, tower/wdec knobs established and restored, real engage witnesses (the mixed lane's tower encode + wdec step deltas). A tensor-QK garble control showed tiny's hs=64 attention does not take that route, so the cell gates the crowned PIPELINE; per-variant kcov witnesses are ledgered as followup 60b.

3. The catalog fixture regeneration rail

capture_catalog.das boots the real setup-mode server on a scratch catalog dir, downloads the smallest card for real (3.1 GB, the server's own sha gate verifying), snapshots /catalog mid-flight and at the terminal state, normalizes machine-local paths, and deletes the scratch. The committed catalog_downloading.json/catalog_done.json are its output - the 7-card roster, and a genuine 5% mid-flight snapshot (the first capture's 32 KB first-chunk pinned 0% and hollowed the progress-render assert - the round's WEAKENED verdict; the rail now waits for real progress). The fixtures' path normalization is pinned by a must-red cell; live-switch.spec synthesizes its second-present-row precondition on a clone (the established withVisionPresent pattern). Playwright 106/106.

The round

Woodpecker (one P1: the reused per-function table is now cleared, never deleted - the documented delete-then-reuse hazard), a worktree tdd audit (its three cheap named gaps closed in-PR), and five checklist auditors. Checklist repairs land across tests (skip-key property, family-tag scoping to model suites, box-decided predicate scope, the lane-knob roster gains whisper's knobs, the stage+mint rule names its APIs with the facade-knob residue ledgered as followup 61), benchmarks (engine-scoped tune_gate, reported-walls subtraction), dasLLAMA (attribution states the property - your own callee or data key is not attribution), and the server (captured-body property form, the playwright-suite duty, transitive-require routing).

Chain green first pass: all preflight lanes, model-free suite (whisper's model-gated cells included under the runner env), test_model_catalog 18/18, playwright 106/106.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LY6gNeDZkF14X81rLz3QLc

[Copilot]

Pull request overview

Follow-up PR that tightens dasLLAMA review/test gates and improves server fixture regeneration to prevent silent regressions in GPU hand-binding and measurement harness logic.

Changes:

• Add/extend review gates and docs around Metal race/knockout hand-bind validation (check_race_bind_numbers) and related GPU review policy.
• Add model-free test coverage for measurement orchestrator “pure seams” and canary bench script self-test hooks.
• Add a scripted fixture capture rail for dasllama-server catalog fixtures and update Playwright tests/fixtures accordingly.

Reviewed changes

Copilot reviewed 21 out of 21 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
utils/dasllama-server/tests/live-switch.spec.js	Synthesizes a second “present” catalog row in tests now that capture rail downloads only one card.
utils/dasllama-server/tests/fixtures/README.md	Documents scripted capture rail for catalog fixtures.
utils/dasllama-server/tests/fixtures/catalog_downloading.json	Updates mid-flight catalog snapshot fixture to new capture rail output.
utils/dasllama-server/tests/fixtures/catalog_done.json	Updates post-download catalog snapshot fixture to new capture rail output.
utils/dasllama-server/tests/fixtures/capture_catalog.das	New scripted fixture regeneration rail that boots server, downloads one real card, snapshots /catalog, normalizes paths, and cleans up scratch dir.
utils/dasllama-server/test_model_catalog.das	Adds a test pinning fixture path normalization (no machine-local path leakage).
utils/dasllama-server/REVIEW.md	Tightens fixture authoring/capture rules and requires Playwright runs when fixtures/control page change.
modules/dasLLAMA/tests/test_whisper.das	Adds crowned word-level equality test with normalization and GPU-engage witnesses.
modules/dasLLAMA/tests/test_gen_records_args.das	New model-free tests for orchestrator seams: pybench args, GPU pair roster, canary self-test, and zero-cells verdict.
modules/dasLLAMA/tests/run.das	Registers the new test_gen_records_args.das in the suite list.
modules/dasLLAMA/tests/REVIEW.md	Clarifies skip-gate rules, lane-pin/loader invariants, family tagging scope, and transcript grading policy.
modules/dasLLAMA/tests/CLAUDE.md	Documents new test coverage and the crowned word-level gate intent.
modules/dasLLAMA/REVIEW.md	Broadens the “no attribution” rule to all checked-in prose/comments owned by the module.
modules/dasLLAMA/REVIEW.das	Adds check_race_bind_numbers and supporting helpers; wires the new gate into main().
modules/dasLLAMA/REVIEW_GPU.md	Aligns hand-binding review duties with what the new machine check can/can’t see.
modules/dasLLAMA/performance/profile_common.das	Extracts asr_gpu_pair_tool and records_run_verdict utilities used by orchestrator/tests.
modules/dasLLAMA/performance/gen_bench_records.das	Uses extracted helpers for GPU-pair logic and run verdict (including zero-cells refusal).
modules/dasLLAMA/followup_general.md	Records followups for per-variant crowned witnesses and facade-knob image-off seam.
modules/dasLLAMA/benchmarks/REVIEW.md	Clarifies tune gate applicability and raw-wall reporting rule wording.
modules/dasLLAMA/benchmarks/asr/canary_qwen_bench.py	Adds torch-free self-test seams, argless refusal, prompt rebuild per rep, and zero-BENCH-row failure.
modules/dasLLAMA/ARCHITECTURE_ENGINE.md	Documents the new check_race_bind_numbers gate and its licensed residue.

---

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

[Copilot]

Pull request overview

Copilot reviewed 21 out of 21 changed files in this pull request and generated no new comments.

Suppressed comments (2)

Previously missed (2) — in code that hasn't changed since the last review.

utils/dasllama-server/tests/live-switch.spec.js:23

• withE4bPresent hard-codes the models directory string even though the catalog fixture already provides models_dir. Deriving the row path from d.models_dir avoids duplicating the convention and prevents the helper drifting if fixtures are re-captured with a different normalized prefix.

    e.path = 'C:\\Users\\user\\.dasllama\\models\\' + e.file;

modules/dasLLAMA/REVIEW.das:264

• rb_scan_classes parses class names by taking ident_prefix immediately after "class ", which mis-identifies declarations like class private Foo : ... as class name private. That can break base-class tracking and binding-number collection, causing check_race_bind_numbers to miss or mis-attribute declared bindings.

            var rest = slice(line, length("class "))
            if (rest |> starts_with("template ")) {
                rest = slice(rest, length("template "))
            }
            cur = ident_prefix(rest)