feat(capture): DeepSeek-V4 mHC activation capture targets

[RhizoNymph]

Adds DeepSeek-V4 manifold-hyperconnection (mHC) activations as capture targets on feat/capture-consumers. Same capture-side mHC work as #181 (which targets feat/integration), ported here so the standalone capture-consumers branch supports capturing mHC hook points.

Hook points

hook	fires at	shape	dtype
mhc_streams_pre_attn	multi-stream residual entering the layer	(n, hc_mult, hidden)	bf16
mhc_streams_pre_mlp	multi-stream residual before the FFN fold	(n, hc_mult, hidden)	bf16
mhc_streams_final	final pre-hc_head streams at the model tail	(n, hc_mult, hidden)	bf16
mhc_attn_post_mix / mhc_ffn_post_mix	attn/FFN post-mix gates	(n, hc_mult)	fp32
mhc_attn_res_mix / mhc_ffn_res_mix	attn/FFN Sinkhorn stream-mixing matrices	(n, hc_mult, hc_mult)	fp32

Plus reuse of the standard pre_attn / post_attn / mlp_in / mlp_out hooks for V4's single-stream (n, hidden) bf16 sublayer in/out. mhc_streams_final is a model-level hook (selector normalized to the tail layer).

Changes

• Hook vocabulary (HookName + mirrored id table) extended with the 7 mhc_* names.
• Per-hook HookSchema(width, dtype, logical_shape) + build_hook_schema(hidden, dtype, hc_mult), replacing the single hidden_size/model_dtype assumption; carried on CaptureContext.hook_schema and built in build_capture_context.
• Manager sizes global buffers + per-key scratch dtype from the schema (fp32 coefficients not downcast); chunk metadata carries row_shape + per-row positions.
• Per-model (schema-driven) hook validation; model-level hooks normalize their layer selector.
• Filesystem consumer/reader: row_shape, per-entry dtype (mixed-dtype packed files), per-row positions + latest_per_position().
• V4 decoder-layer + model-tail taps (nvidia + amd, fused-CUDA + native), gated to constant-fold out when capture is disabled.

Backward compatible: additive sidecar fields, unchanged request/CLI surface and .bin payload, standard-model validation unchanged.

Validation

Full capture suite green on this branch (335 passed, 1 skipped) — mHC tests plus the existing prefix-cache/admission tests. The same work was GPU-validated end-to-end on DeepSeek-V4-Flash-NVFP4 / 4× B300 (see #177/#181).