Bump MuTap to M3: nn_suppressor on tap::dsp::nn, pitch search in Sample

[tap]

What this changes

submodules/MuTap → tap/MuTap main at 96b4572 (tap/MuTap#46), which carries DspTap 58210ae (tap/DspTap#15) as its nested pin. No source change in the externals.

Why

The wake-word plan's M3 pass criterion: MuTap and MuTap-Max pins bumped once the learned kernels live in DspTap's tap::dsp::nn and nn_suppressor consumes them. This is the MuTap-Max half. The same MuTap commit also moves the speech predictor's pitch search to accumulate in Sample, the root-cause fix for the Cortex-M33 finding; mutap.aec~ instantiates the double profile, which is bit-for-bit unchanged by that.

Verification

• The pinned MuTap tree merged green on all thirteen of its CI jobs; the learned post engine mutap.aec~ instantiates (aec_chain_nn<double>) was shown bit-identical on the shipping v2 model against the pre-refactor binary, and the six Python parity cases pass at the M2 depths.
• The nested DspTap pin resolves (git submodule update --init --recursive checks out 58210ae, the squash of Add tap::dsp::nn: dense and GRU inference kernels DspTap#15).
• Nothing in the externals reads a suppressor's weight arrays back after construction (the one API nuance of the refactor: nn_suppressor::config::weights is moved into the kernels at construction).
• Not built here (no Max SDK on this Linux container): CI's macOS and Windows external builds are the first real gate for this pin.

Notes for the reviewer

• Submodule pin moved. submodules/MuTap 1d258d3 → 96b4572, the merged tree on MuTap main, so it stays reachable after branch cleanup. Nothing else.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JuUg1ZBxm3fyBqWgQv6H1G

---

Generated by Claude Code