dusterbloom: feat(qwen3_next): mixed-bit Qwen3.5 GDN BA loading fallback

[dusterbloom]

Summary

Adds a load-time fallback for Qwen3.5 models with mixed-bit GDN BA projections (some layers q4, some q8 — common in unsloth dynamic-quant variants). The default fused loader concatenates in_proj_a + in_proj_b into a single matmul; mixed-bit weights have incompatible shapes and fusion fails. This PR detects the specific failure and retries with separate (4-dispatch) GDN projections.

This is PR-1c of the magic-canvas split — a deferred follow-up from PR-1's audit (commit 061e500c).

Behaviour

┌─ load_qwen3_5_model / load_qwen3_5_moe_model ─────┐
│                                                   │
│  args.use_separate_gdn_projections                │
│  || HIGGS_SEPARATE_GDN_PROJ env                   │
│      ↓ yes                                         │
│  ┌── direct (separate, 4 dispatches) ──┐         │
│      ↓ no                                          │
│  ┌── try fused (2 dispatches) ──────────┐         │
│      ↓ ShapeMismatch matching "in_proj_ba" +     │
│        "requires separate GDN projections"        │
│      ↓ retry with use_separate=true              │
│  ┌── direct (separate, mixed-bit fallback) ──┐   │
└───────────────────────────────────────────────────┘

What's added

• is_mixed_bit_gdn_ba_fusion_error(err: &ModelError) -> bool — pattern-matches the diagnostic shape-mismatch error
• qwen3_5_quantization_config — parses {group_size, bits} from the per-layer quantization map
• qwen3_5_mixed_ba_quantization_layers — scans for layers where in_proj_a and in_proj_b differ in bits or group_size
• can_concatenate_axis0 — guard used inside load_qwen3_5_moe_weights_fused to emit the diagnostic error rather than panic on the concat
• load_qwen3_5_model_with_gdn_fallback — private helper called by both dense and MoE load paths, unifies the fallback

Adaptations from feat/magic-canvas

• The dense load_qwen3_5_model previously only honoured the env var; now it honours args.use_separate_gdn_projections too, matching the MoE path. Strict improvement.
• Direct cherry-pick of 061e500c had 5 conflict regions; this is a manual surgical port that preserves origin/main's structure.

Hygiene

• No unwrap() on Result/Option
• No as casts in shape-arithmetic paths (use i32::try_from)
• Match arms enumerate variants
• No file-level blanket allows added

Test plan

• cargo check -p higgs-models — clean
• cargo clippy --all-targets --all-features -- -D warnings — clean (rustc 1.95.0, matches CI)
• cargo fmt --check — clean
• cargo test -p higgs-models --lib — 333/333 pass (3 new)
• End-to-end: load an unsloth dynamic-quant Qwen3.5 (e.g. Brooooooklyn/Qwen3.5-27B-unsloth-mlx) and confirm the warn-then-retry path triggers, model loads successfully, generation produces coherent tokens.

🤖 Generated with Claude Code