Adopt the DspTap FIR substrate via submodule

[tap]

The adoption half of M0 in the RatioTap plan (Appendix B), following tap/DspTap#5: SampleRateTap now consumes the shared design/format/kernel/measurement layer from DspTap instead of carrying it privately. Net −820 lines. The MuTap fft.h re-export shim is the pattern throughout, so every historical include path and name keeps compiling.

What changed

• Submodule: submodules/dsptap pinned at the substrate-extraction merge (98abb04); tap::dsp linked into the SampleRateTap::SampleRateTap INTERFACE target. All workflow checkouts now use submodules: recursive (the MuTap convention); the SessionStart hook already ran submodule init.
• srt/detail/kaiser.h → re-export shim: tap::samplerate::detail names unchanged.
• srt/sample_traits.h → refinement: each specialization derives from the tap::dsp format core and adds the ASRC-specific blend stratum (make_blend_factor, blend_factor_from_q64, blend — the mu-interpolation machinery a fixed-ratio converter has no use for, which is why it stays here). The sample_type concept states the full combined contract; detail::round_sat/clamp_sat re-exported.
• srt/polyphase_filter.h → dot kernels consumed from tap::dsp via using-declarations (SRT_* macro names forwarded to the TAP_DSP_* gates, including the SRT_CP_MIN_CHANNELS override); the bank constructor's inline largest-remainder correction replaced by tap::dsp::quantize_row_preserving_sum. The bank, mu-blend functions, and fractional_resampler are untouched.
• tests/support/ analysis headers → shims over tap::dsp::analysis; the srt_test names keep working.
• Book: 21 anchor includes repointed to the submodule paths (Kaiser design, dot kernels, row-sum quantization, the moved sample-traits regions, the multitone/sine instruments). Every remaining {{#include}} verified to resolve against its anchors — the book CI's stale-anchor gate re-verifies. The sample-traits and optimization chapters now show the refinement/relocated code where their anchors moved; a prose pass over those chapters is deliberately out of scope here so the diff stays mechanical.

Behavior change: none — and how that's proven

• Full test suite green under GCC and clang -Werror (73/73 both — includes the SNR quality suites, DC-gain exactness, and multichannel bit-exactness tests, which would catch any arithmetic drift).
• scripts/tidy.sh clean over the affected TUs; clang-format clean.
• The icount CI gates (±3% on Cortex-M33/M55/Hexagon) run on this PR and are the strongest proof the relocation costs nothing on the hot path: identical kernels, now included from the submodule, must land inside the committed baselines.

One reviewable nuance: the bank constructor now accumulates each row's exact sum in storage (tap-reversed) order rather than prototype order inside the shared utility. The per-tap quantization and largest-remainder scan are order-identical; only the llround input can differ in the last ulp, which the DC-gain and row-sum tests bound to the same one-LSB guarantee as before.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Ldeq57sBySx2nFTQsGcQB6

---

Generated by Claude Code