Big, clean change. The SMI removal is thorough — collector, the ~400-line XML model, and the 4.2k-line fixture all gone — and the four-commit structure makes it reviewable. Driver-free PCI discovery with GFD-label precedence is a real day-0 improvement, and dropping the CUDA base is a solid CVE-surface win. The precedence tests (TestFromMeasurements_PCIBackfill: supported->backfill, unsupported->GPUModel-only, label-wins-over-PCI) are the right matrix.

Code review

• .goreleaser.yaml: removing aicrd's base_image looks like a side effect past the stated scope — aicrd moves from gcr.io/distroless/static:nonroot to ko's chainguard default. Confirm it's intended and that the nonroot UID posture survives. Inline.
• nfd.go: the new heterogeneous / multi-SKU resolution in extractHardwareInfo has no direct unit test (no nfd_test.go change in the PR). Inline.
• from_measurements.go: a day-0 unsupported SKU now leaves Accelerator empty with no note, where the old smi path set unknown-sku. Looks deliberate (you dropped the old test), but the rollout note says "accelerator: unknown" — worth reconciling. Inline.
• device_ids.go: the table is only as correct as the pci.ids curation; the positive tests can't catch a mis-sourced id. Inline.

Go review

• skuForDeviceID: the 0x strip is case-sensitive and runs before ToLower, so a 0X-prefixed id slips through. Minor, but easy to make uniform. Inline.
• Error and degradation paths read well — Collect returns no subtypes rather than erroring, the GetString early-returns are consistent, and no concurrency surface is introduced.

Non-blocking: this is stacked on #1350 (base fix/1349-..., not main), so re-target once #1350 lands. Review order is #1350 first, then this.

Re-reviewed at 42f1f15 — this commit cleanly addresses all five points from the last pass, and I traced each:

• aicrd base image (.goreleaser.yaml): resolved. The comment documents that aicrd intentionally rides ko's chainguard/static default and stays nonroot UID 65532, matching the old distroless/static:nonroot. That's correct for ko's default user. Only nice-to-have: assert the nonroot user on the built image in CI so the posture can't silently regress later — not a blocker.
• Heterogeneous PCI path (nfd_test.go): resolved. The new TestExtractHardwareInfo cases cover the three I flagged — single SKU resolves, two distinct SKUs -> "", known-among-unknown -> the known SKU — with real wantSKU assertions that go red if the guard breaks.
• unknown-sku note (from_measurements.go): resolved, and the precedence holds up. A day-0 unsupported SKU now records Accelerator{Note: unknown-sku} with the descriptive value in GPUModel; supported SKUs still backfill the value; a label-resolved value or multi-gpu note is left untouched; and a topology-set unknown-sku note isn't clobbered by the PCI source. The early-return refactor also reads cleaner than the old compound condition, and matches the smi behavior the rollout notes describe.
• 0x prefix (device_ids.go): resolved. ToLower runs before TrimPrefix now, and the new 0X2330 test case discriminates against the old ordering.

One standing, non-blocking item: the device-ID table's positive entries are still only as trustworthy as the pci.ids curation (the tests pin them but can't catch a mis-sourced id). Nothing for this commit to change — leaving it as a spot-check / generator suggestion if the table grows.

No blockers from the code. CI is green except the two H100 GPU e2e jobs still running; the full Merge Gate suite will run once this retargets to main after #1350 lands. Stacked-on-#1350 merge order still stands.

Approving as an aicr-maintainer code owner. Re-reviewed at 42f1f15 — all five points from the prior pass are addressed and verified.

This satisfies the code-owner review for the maintainer-owned files in this PR (.goreleaser.yaml, .settings.yaml, .github/**). The Go code, docs, and examples are under the default aicr-write owner and still need an approval from that team.

Heads-up: this approval will likely need re-doing once #1350 lands and this branch rebases/retargets onto main (new commit SHAs will mark it stale).

Re-approving after the retarget and rebase onto main (the prior approval was dismissed when the SHA changed). I verified the rebased tree at 11f2004 matches what I approved: same 32-file change, gpu_sku.go correctly dropped (now in main via #1350), and all five review fixes intact — the 0x prefix ordering, the unknown-sku note on the PCI path, the nonroot-65532 note on the aicrd base, and the new heterogeneous and known-among-unknown nfd tests.

This satisfies the aicr-maintainer code-owner review for the maintainer-owned files (.goreleaser.yaml, .settings.yaml, .github). The Go code, docs, and examples are under aicr-write and still need an approval from that team.

Note: the full Merge Gate is running against main for the first time on this branch; the merge will correctly hold until those checks are green.

Actionable comments posted: 2

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@docs/integrator/data-flow.md`:
- Line 190: The GPU documentation contains mixed guidance showing both outdated
SMI/driver-era examples and the new driver-free hardware/model paths, which will
confuse readers about which snapshot paths are currently supported. At
docs/integrator/data-flow.md line 190, update the downstream recipe-structure
example to remove any references to driver/cudaVersion and instead use
hardware.model to align with the driver-free architecture. At
docs/integrator/recipe-development.md line 313, replace the existing
check-nvidia-smi or driver-version based example with a corresponding
hardware/model path example. Ensure both locations consistently reflect that the
snapshot system is now driver-free and hardware/model-based, removing all stale
SMI-era GPU field references.

In `@pkg/collector/gpu/hardware.go`:
- Around line 44-48: The docstring for the SKU field in the HardwareInfo struct
incorrectly describes it as containing only "AICR accelerator enum values,"
which misrepresents its actual contract. Update the docstring to clarify that
SKU contains GPU SKU identifiers (such as "h100", "l40") resolved from the GPU's
PCI device ID mapping, which represents a broader descriptive vocabulary than
just recipe-supported enum values. Preserve the existing details about when SKU
is empty (unknown device ID or heterogeneous mix) and its use in fingerprinting
without nvidia-smi or GFD node labels.

---

Outside diff comments:
In @.github/actions/aicr-build/action.yml:
- Around line 23-25: Update the description for the build_snapshot_agent input
in the action.yml file to accurately reflect that the action now supports
multiple types of snapshot agent images (driver-free/static base images) rather
than only CUDA-based images. Replace the text "CUDA-based snapshot agent image"
with more inclusive language that covers all supported image types while
maintaining the mention of loading it into kind.

In `@docs/integrator/validator-extension.md`:
- Around line 185-233: The catalog entry name and description do not match what
the check script actually validates. The check script validates the GPU
SKU/model against an allow-list, but the catalog entry is named
gpu-driver-version with a description about driver version validation. Update
the catalog entry's name field to reflect GPU model validation (such as
gpu-model-check or gpu-sku-validation) and update the description field to
clearly state it verifies the GPU accelerator model/SKU against an allowed list,
removing any reference to driver version.