E2: V3.2 DSA indexer (MODEL_DEEPSEEK_V32)

[tlkahn]

V3.2 DSA indexer (MODEL_DEEPSEEK_V32)

Port mlx-lm deepseek_v32.py Indexer - a real attention/cache fork, not a boolean on collapsed V4.

Spec: doc/plans/engine-inference-core.md § E2

Part of: #105 (E2) · Epic: #56
Depends on: MLA attention + MLA KV cache; family split (MODEL_DEEPSEEK_V32)

Tasks

• Indexer module: top-k index heads (index_n_heads / index_head_dim / index_topk)
• Index-side RoPE + index K cache
• Sparse gather (take_along_axis) on decode
• Sparse mask on prefill
• Dual cache entry (MLA latent cache + indexer K cache), matching mlx-lm CacheList
• Wire indexer weight names through load/sanitize (wq_b, wk, weights_proj, ...)
• Config already parsed: index_head_dim, index_n_heads, index_topk (from family-split/config issue)

Notes

• Decision 11: this fork is why V3.2 must not share MODEL_DEEPSEEK_V4 / V3 tags