schema: implement gpu_hardware table with auto-detection

[AlbinoGeek]

Parent: #8 — Additional Performance Metrics
Depends on: #17 (interface_type table)

Implement the gpu_hardware entity table as ratified in #8. This table stores die-level GPU intrinsics — fixed properties of the GPU model, not of a specific system configuration.

Schema

CREATE TABLE gpu_hardware (
    gpu_hardware_id              SERIAL PRIMARY KEY,
    gpu_name                     TEXT NOT NULL,
    pci_device_id                TEXT NULLABLE,           -- "0x2684"
    pci_sub_device_id            TEXT NULLABLE,           -- board partner variant
    vram_total_mb                INT NULLABLE,
    vram_type                    TEXT NULLABLE,           -- "GDDR6X", "HBM2e", etc.
    memory_bus_width_bits        INT NULLABLE,
    memory_bandwidth_peak_gb_s   FLOAT NULLABLE,          -- stored (not derived) per design decision
    clock_memory_mhz             INT NULLABLE,
    clock_graphics_boost_mhz    INT NULLABLE,
    peak_tflops_fp16             FLOAT NULLABLE,
    tdp_watts                    INT NULLABLE,
    gpu_native_interface_type_id INT NULLABLE REFERENCES interface_type  -- card's rated spec
);

Auto-detection requirements

The harness should auto-detect and populate a gpu_hardware row from the running system. Detection sources:

• gpu_name — nvidia-smi --query-gpu=name --format=csv,noheader
• pci_device_id / pci_sub_device_id — nvidia-smi --query-gpu=pci.device_id,pci.sub_device_id
• vram_total_mb — nvidia-smi --query-gpu=memory.total (convert MiB → MB)
• vram_type — PCI device ID lookup table (no direct nvidia-smi query; deterministic from device ID)
• memory_bus_width_bits — pynvml.nvmlDeviceGetMemoryBusWidth()
• clock_memory_mhz / clock_graphics_boost_mhz — nvidia-smi --query-gpu=clocks.max.memory,clocks.boost.graphics
• peak_tflops_fp16 — existing _TFLOPS_TABLE lookup (already implemented in harness); None if not in table
• memory_bandwidth_peak_gb_s — computed from clock_memory_mhz, memory_bus_width_bits, and vram_type (DDR factor)
• gpu_native_interface_type_id — FK to interface_type row matching the GPU's rated PCIe spec

Identity / deduplication

A gpu_hardware row is a model-level record, not an instance record. Two users with the same GPU model should share a single row. Deduplicate on (pci_device_id, pci_sub_device_id) when both are available; fall back to gpu_name normalization.

Fabrication detection: pci_device_id must be consistent with the gpu_name reported — cross-check before insert. Flag mismatch as invalid.

Acceptance criteria

• Migration creates gpu_hardware with all columns and FK to interface_type
• Harness auto-detects and upserts a gpu_hardware row at startup (nvidia-smi + pynvml)
• Deduplication logic: existing row with matching PCI IDs is reused, not duplicated
• memory_bandwidth_peak_gb_s is computed and stored at detection time
• peak_tflops_fp16 lookup uses existing _TFLOPS_TABLE
• Migration is reversible

— Bastion