Root cause analysis for device-bound firmware leaking into hardware.firmware
on Supermicro Redfish (SYS-A21GE-NBRT HGX B200):
- collectFirmwareInventory (6c19a58) had no coverage for Supermicro naming.
isDeviceBoundFirmwareName checked "gpu " / "nic " (space-terminated) while
Supermicro uses "GPU1 System Slot0" / "NIC1 System Slot0 ..." (digit suffix).
- 9c5512d added _fw_gpu_ / _fw_nvswitch_ / _inforom_gpu_ patterns to fix HGX,
but checked DeviceName which contains "Software Inventory" (from Redfish Name),
not the firmware Id. Dead code from day one.
09-testing.md: add firmware filter worked example and rule #4 — verify the
filter checks the field that the collector actually populates.
10-decisions.md: ADL-019 — isDeviceBoundFirmwareName must be extended per
vendor with a test case per vendor format before shipping.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
On Supermicro HGX systems (SYS-A21GE-NBRT) ~35 sub-chassis (GPU, NVSwitch,
PCIeRetimer, ERoT/IRoT, BMC, FPGA) all carry ChassisType=Module/Component/Zone
and expose empty /Drives collections. shouldAdaptiveNVMeProbe returned true for
all of them, triggering 35 × 384 = 13 440 HTTP requests → ~22 min wasted per
collection (more than half of total 35 min collection time).
Fix: chassisTypeCanHaveNVMe returns false for Module, Component, Zone. The
candidate selection loop in collectRawRedfishTree now checks the parent chassis
doc before adding a /Drives path to the probe list. Enclosure (NVMe backplane),
RackMount, and unknown types are unaffected.
Tests:
- TestChassisTypeCanHaveNVMe: table-driven, covers excluded and storage-capable types
- TestNVMePostProbeSkipsNonStorageChassis: topology integration, GPU chassis +
backplane with empty /Drives → exactly 1 candidate selected (backplane only)
Docs:
- ADL-018 in bible-local/10-decisions.md
- Candidate-selection test matrix in bible-local/09-testing.md
- SYS-A21GE-NBRT baseline row in docs/test_server_collection_memory.md
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Three bugs, all related to GPU dedup in the Redfish replay pipeline:
1. collectGPUsFromProcessors (redfish_replay.go): GPU-type Processor entries
(Systems/HGX_Baseboard_0/Processors/GPU_SXM_N) were not deduplicated against
existing PCIeDevice GPUs on Supermicro HGX. The chassis-ID lookup keyed on
processor Id ("GPU_SXM_1") but the chassis is named "HGX_GPU_SXM_1" — lookup
returned nothing, serial stayed empty, UUID was unseen → 8 duplicate GPU rows.
Fix: read SerialNumber directly from the Processor doc first; chassis lookup
is now a fallback override (as it was designed for MSI).
2. looksLikeGPU (redfish.go): NVSwitch PCIe devices (Model="NVSwitch",
Manufacturer="NVIDIA") were classified as GPUs because "nvidia" matched the
GPU hint list. Fix: early return false when Model contains "nvswitch".
3. gpuDocDedupKey (redfish.go): commit 9df29b1 changed the dedup key to prefer
slot|model before path, which collapsed two distinct GPUs with identical model
names in GraphicsControllers into one entry. Fix: only serial and BDF are used
as cross-path stable dedup keys; fall back to Redfish path when neither is
present. This also restores TestReplayCollectGPUs_DedupUsesRedfishPathBeforeHeuristics
which had been broken on main since 9df29b1.
Added tests:
- TestCollectGPUsFromProcessors_SupermicroHGX: Processor GPU dedup when
chassis-ID naming convention does not match processor Id
- TestReplayCollectGPUs_DedupCrossChassisSerial: same GPU via two Chassis
PCIeDevice trees with matching serials → collapsed to one
- TestLooksLikeGPU_NVSwitchExcluded: NVSwitch is not a GPU
Added rule to bible-local/09-testing.md: dedup/filter/classify functions must
cover true-positive, true-negative, and the vendor counter-case axes.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add bible.git as submodule at bible/
- Move docs/bible/ → bible-local/ (project-specific architecture)
- Update CLAUDE.md to reference both bible/ and bible-local/
- Add AGENTS.md for Codex with same structure
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
