feat(inference): CBAR-PIECE-5 PR-2 — GGUF metadata loader populates QwenModelMetadata

[joelteply]

Summary

Stacks on #1331 (CBAR-PIECE-5 PR-1, residency gate types). PR-1 defined the typed gate; this PR-2 reads a real GGUF file and produces the QwenModelMetadata the gate consumes. PR-3 wires both probe + this loader into the turn dispatcher with enforcement.

Same pure-functions cadence as PR-1 — file I/O lives in a thin wrapper, all parsing logic lives in helpers unit-testable without GGUF fixtures.

What ships in inference_capability/gguf_loader.rs

• pub fn read_qwen_model_metadata(path: &Path) -> Result<QwenModelMetadata, String> — thin file-opener using backends::gguf_file::Content already in the crate. No new dependencies.
• pub(crate) fn file_type_to_bytes_per_param(ft: u32) -> Result<f64, String> — maps GGUF general.file_type enum to bytes-per-weight. Covers the full shipped quantization set (Q4_0/Q4_1/Q4_K_S/Q4_K_M/Q5_0/Q5_1/Q5_K_S/Q5_K_M/Q6_K/Q8_0, IQ-series sub-2-bit, F16/F32/BF16). Unknown ft returns Err with the value named.
• pub(crate) fn layer_kinds_for_architecture(arch: &str) -> Vec<String> — lookup table for Vulkan-llama.cpp gaps: qwen3moe → [moe_gate, sliding_window_attn], qwen3 → [sliding_window_attn], everything else → []. Pinned by a dedicated test so renames must land in both the table + residency.rs's matching test simultaneously.

Failure-mode discipline

Field	Required?	Action when missing
general.architecture	✅	Err with named path
{arch}.block_count	✅	Err with named path
general.file_type	✅	Err with named path
general.parameter_count	optional	derive from file_size / bytes_per_param (loud fallback, documented in field comment)
general.name	optional	fall back to file stem (display only)

Same posture as backends::read_gguf_metadata (Joel's 2026-04-23 fix removed all silent-llama-fallback paths there).

Test plan

15 passing on cargo test --lib --features metal,accelerate inference_capability::gguf_loader::

• file_type_to_bytes_per_param (7):
  • workhorse quants (0, 1, 2, 7, 8, 14, 15, 17, 18, 32) all have entries
  • Q4_K_M in 0.55-0.65 bytes/param band
  • FP16 = 2.0 exactly
  • F32 = 4.0 exactly
  • unknown ft (9999) returns Err with value named
  • removed ft (4, 5, 6) return Err
  • quantization ordering monotone (Q4_K_M < Q5_K_M < Q6_K < Q8_0 < F16 < F32)
  • IQ-series sub-2-bit all < 0.4 bytes/param
• layer_kinds_for_architecture (5):
  • qwen3moe lists both moe_gate + sliding_window_attn
  • qwen3 lists sliding_window_attn only
  • qwen2 + qwen2vl empty (Vulkan handles them)
  • unknown arch (mistral, phi3, empty, future-model) returns empty
  • table-pinning test forces rename to land in two places
• read_qwen_model_metadata I/O (2):
  • nonexistent path returns Err with descriptive message
  • non-GGUF file (Cargo.toml) returns Err

VDD evidence

N/A — pure-data loader, no inference dispatch. Evidence lands with PR-3 (enforcement integration).

Stack

• feat(grid-inference-routing): PR-1 InferenceCapability + probe + NodeCapabilityRegistry #1315 GRID-INFERENCE-ROUTING PR-1 — merged to canary
• feat(inference): CBAR-PIECE-5 PR-1 — Qwen GPU residency gate (pure-functions slice) #1331 CBAR-PIECE-5 PR-1 (residency gate types) — base of this PR, OPEN
• This PR: GGUF metadata loader (CBAR-PIECE-5 PR-2)
• Future PR-3: dispatcher integration + enforcement (block the turn instead of letting llama.cpp silently split layers to CPU)

After #1331 merges, this PR rebases cleanly onto canary.