Summary

The GPU SKU classifier (ParseGPUSKU) identifies a node's accelerator by substring-matching nvidia-smi's ProductName against a pattern table. Because it uses strings.Contains, any SKU whose product name contains a shorter registered pattern is silently collapsed onto the wrong accelerator. This defeats the classifier's own unknown-sku safety signal and produces fingerprints that misrepresent the hardware.

• Classifier: pkg/fingerprint/gpu_sku.go#L48
• Pattern table: pkg/fingerprint/gpu_sku.go#L24-L41
• unknown-sku fallback this defeats: pkg/fingerprint/from_measurements.go#L186-L194

Current behavior / evidence

On an Oracle OKE cluster with nvidia.com/gpu.product=NVIDIA-L40S, the snapshot fingerprint records:

fingerprint:
  accelerator:
    source: nodeTopology.label.nvidia.com/gpu.product
    value: l40        # the trailing "S" is dropped

NVIDIA-L40S contains the substring L40, so it matches {"L40", "l40"} and is reported as a (valid-but-wrong) l40 instead of being flagged unknown-sku.

This is a class of bug, not a one-off. The table already carries two hand-patched workarounds for the same root cause:

• {"GB200", ...} is ordered before {"B200", ...} so a Grace-Blackwell node isn't read as b200.
• {"GH200", ""} is an explicit exclusion so Grace-Hopper isn't read as h200.

Each new collision currently requires another manual ordering/exclusion entry. Known/likely live collisions that are not yet patched:

• NVIDIA RTX A1000 → "A1000" contains "A100" → silently fingerprints as datacenter a100.
• NVIDIA L40S → l40 (the reported case).

Proposed approach

Replace substring matching with token-boundary matching in ParseGPUSKU: tokenize the product name (split on whitespace and -) and match whole tokens, with contiguous-token-sequence matching for multi-word patterns (RTX PRO 6000). Unrecognized tokens fall through to unknown-sku (the honest signal).

This structurally eliminates the entire collision class and removes the need for per-SKU ordering/exclusion hacks:

• NVIDIA L40S → token L40S (not L40) → no false match → unknown-sku.
• NVIDIA RTX A1000 → token A1000 (not A100) → no false a100.
• NVIDIA A100-SXM4-40GB → token A100 → a100. ✓
• NVIDIA GB200 NVL72 → token GB200 → gb200; B200 is no longer a substring concern, so the GB200-before-B200 ordering hack and the GH200 exclusion become unnecessary.

Acceptance criteria

• ParseGPUSKU matches on token boundaries (handles space and - delimiters; supports the multi-word RTX PRO 6000 pattern).
• Unrecognized product strings return "" so the fingerprint records unknown-sku (raw model already preserved in measurements).
• Ordering/exclusion workarounds removed where token matching makes them redundant (GB200/B200, GH200), with tests proving the behavior still holds.
• Test matrix asserts decoys do not false-map: RTX A1000, L40S, L4, A800, H800, GH200, plus all currently supported SKUs continue to resolve correctly.

Notes / related

• Making l40s (or any newly distinguished SKU) a first-class, selectable accelerator is a separate, larger change (criteria enum, OpenAPI, CLI/API, docs) and should land together with actual recipe coverage for that SKU — otherwise selecting it just yields "no matching recipe." This issue is only about fixing the misclassification, not adding new enum values.