[cacheline999]

Governing Issue and delivery slice

Refs #38

• Accepted scope / delegation: task: add MLX backend discovery and seed source-grounded MLX knowledge for Apple Silicon #38 Phase 1 — "unit tests cover the Apple, non-Apple and model-mismatch paths" and "an optional case-insensitive --device-model match". Not yet assigned; real-host evidence and this scope are posted on task: add MLX backend discovery and seed source-grounded MLX knowledge for Apple Silicon #38 (task: add MLX backend discovery and seed source-grounded MLX knowledge for Apple Silicon #38 (comment)) for the acceptance owner.
• What this PR delivers: whole-name --device-model matching for the MLX probe, mocked tests for the Darwin branches of inspect_devices, and a doctor test that runs only on a real Apple Silicon host.
• What it does not deliver: Intel Mac validation, any MLX/Metal compatibility, model-support or performance claim, changes to the CUDA/ROCm/Ascend matching rules.

Knowledge authorship declaration

Not applicable: no path protected by docs/knowledge-model-policy.md is changed.

Changed artifacts and evidence

Requirement	Actual change	Check / evidence and result	Remaining gap
Case-insensitive --device-model match that verifies the right chip	src/inferrsi/infra/platforms/mlx.py: compare the whitespace-normalized, case-folded request with the whole chip name, with or without the Apple prefix	Before: on a real M1 Pro, --device-model M1, pro and Apple M all returned available / selection_verified: true. After: those three exit 1 with device_model_mismatch; m1 pro still exits 0 with available	none
Unit tests for Apple, non-Apple and model-mismatch paths	tests/test_mlx_discovery.py: mocked Darwin cases (available, model match, partial/other model mismatch, non-ARM64 chip, failed sysctl), labeled as claiming no hardware validation	The mismatch cases for M1, pro, Apple M fail on main and pass with this change	none
Real Apple Silicon run	test_actual_apple_silicon_doctor_reports_available, skipped unless Darwin/arm64	Passed on Apple M1 Pro, macOS 26.1	Intel Mac negative path not run

Checks and dependencies

• Checks actually run (result and tested scope): pytest tests/test_mlx_discovery.py tests/test_inspection.py → 34 passed, 1 skipped (Python 3.11.14, Apple M1 Pro, macOS 26.1); ruff check src tests tools → passed; git diff --check → clean.
• Failed, skipped, timed-out or not-run checks (if applicable): test_actual_non_macos_doctor_reports_unavailable skips on macOS by design; full suite not run locally.
• Dependencies / shared interfaces / integration assumptions (if applicable): inspect_devices signature and result keys are unchanged.

Remaining work and Issue closure

• Remaining Issue requirements / follow-up: Intel Mac negative path; the platform_packs criterion of task: add MLX backend discovery and seed source-grounded MLX knowledge for Apple Silicon #38 is obsolete after refactor: remove the unread platform_packs descriptors and fix the generated-path guidance #201.
• Goal-acceptance owner or accepted closure delegation: task: add MLX backend discovery and seed source-grounded MLX knowledge for Apple Silicon #38 acceptance owner.
• Goal ready for acceptance? Not claimed.

🤖 Generated with Claude Code

[Dong1017]

The whole-name MLX selection fix looks correct at 4980bf1c; hosted Ready validation is still pending.

• The two-file change rejects partial names such as M1, pro and Apple M for Apple M1 Pro, while preserving case/whitespace normalization and the optional Apple prefix. The discovery result contract and other backends are unchanged.
• The applicable Ready run is action_required with zero jobs. Passing declaration checks do not replace its tests/package checks; the next step is the existing maintainer's decision on that run.
• The existing author, cacheline999, retains the correction. The M1 Pro receipt is author-reported; Intel Mac behavior and the parent Issue's acceptance remain separate. Mocked Darwin tests establish software-branch coverage, not hardware compatibility.

Evidence and review scope

• Head: 4980bf1c80e5d4ec134c72ea060499f062bc93e9; API base / observed main: 2baebf6ebb815ebf07d21c41e3b3754858c0798e.
• Independently read both changed files and followed inspect_device through the MLX adapter and environment-inspection caller. The matching predicate now compares normalized whole names rather than substrings.
• Issue evidence distinguishes the author's historical 28d677ca discovery result from this fix. The PR reports targeted tests and Ruff on M1 Pro; I did not reproduce those runs.
• Static review only: no local suite, Apple host, model/device execution or workflow action by this maintenance session. No source defect found in this bounded delta. This COMMENT does not imply APPROVE/LGTM, parent-Issue acceptance or MLX/Metal performance validation.

[amy-why-3459]

LGMT

PR review verdict — PASS
Reviewed base 663f3b3 and head 4980bf1.

• src/inferrsi/infra/platforms/mlx.py: --device-model now matches the whitespace-normalized, case-folded whole chip name (with or without the Apple prefix), so M1 / pro / Apple M no longer select Apple M1 Pro; result keys and the inspect_devices signature are unchanged.
• tests/test_mlx_discovery.py: mocked Darwin cases cover available, whole-name match, partial/other-model mismatch, non-ARM64 and failed sysctl; real-host test is skipped off Apple Silicon.
• Locally ran pytest tests/test_mlx_discovery.py tests/test_inspection.py on Linux: 34 passed, 1 skipped.

Not verified: Apple hardware run (author-reported) and the hosted Ready run, which is still awaiting maintainer approval.

[Dong1017]

Incremental review — approval-record contract (current head 4980bf1c, no new code finding): The November 4 approval starts with LGMT, while the repository's review convention requires the exact, case-sensitive LGTM on the first non-empty line for a qualifying PASS approval. GitHub's APPROVED state alone does not establish that the review meets this additional convention; the main ruleset also requires two eligible, current approvals.

Suggested minimal correction: ask the original independent reviewer to submit a qualifying review with the exact first line (or obtain another eligible independent review). Preserve the existing review record rather than silently treating its typo as compliant. This is a process/evidence finding only: the two-file MLX patch and its tests were already covered in earlier review, and I have not run any tests, altered approvals, or verified Apple hardware or Ready CI.