Bump dsptap to the srdif engine; close the convolution engine's IR-load race; gate FFT sizes

[tap]

What this changes

submodules/dsptap moves from 41215c6 to 2137d86 (DspTap main, the clean-room srdif engine, header-only). conv_engine gets its own scratch buffer for load_ir() plus a reader guard on the IR slot swap, conv_engine::configure and stft::configure (behind spectra and nr) refuse FFT sizes the engine does not support, and the three notebooks that run through the real FFT are re-executed.

Why

Found by an adversarial audit of the FFT and its consumers (2026-10-01):

• Race. load_ir() on the control thread and process_block() on the audio thread transformed through one scratch buffer, so an IR load during playback corrupted either the partition being built or the audio frame. A standalone repro against the old header measured ~5e5 corrupted samples per run of 3.3e5; the new header measures zero. The double-buffer slot swap also had an ABA window (a block reads the slot it claimed at its start; a second load inside that block wrote it), closed by the audio thread publishing the slot it reads and load_ir waiting for it (bounded; refused on timeout, never written while in use). The FFT object stays shared, which DspTap's k_is_shareable permits and a static_assert pins.
• Size gates. DspTap's constructor checks nothing in a release build: measured under ASan, N = 12, 24, 96 and 3072 overflow the heap and N = 1000 returns garbage silently. Nothing in TapTools or its C ABI gated the size. configure now returns false and leaves the object unchanged; process() emits silence until a configure succeeds, since it indexed unallocated buffers otherwise. The C ABI returns -1 on refusal (the Python bridge already raises on it).
• Install rule. install(TARGETS … tap_dsp_fft) failed every top-level configure against the new pin, where that target exists only under TAP_DSP_FFT_CMSIS.

Verification

• Release GCC 13: 325 / 325 Catch2 cases (three new scenarios: the gate, unconfigured silence, and the concurrent IR load, which fails on the old engine and passes here).
• scripts/tidy.sh on the changed TUs: no finding in the changed files (the 46 in taptools_capi.cpp are identical on main). clang-format applied.
• Notebooks re-executed through the C ABI: convolution_reverb exactness against direct convolution tightened from 3.13e-12 to 1.14e-13 and block-size independence from 1.81e-12 to 7.11e-14; every printed number in harmonizer and tune is unchanged. The book cites none of the moved numbers.
• Not verified here: TapTools-Max. configure now returns a bool its wrappers may want to check, and load_ir returns false on the refusal path; the Max attribute setters that decide whether a non-power-of-two size ever reaches stft are in that repo.

Notes for the reviewer

• Submodule pin moved to 2137d86, DspTap main. The engine swap moves every float and double output bit (the tests tolerate it; the notebooks are re-executed for it).
• Notebooks re-executed, because behavior changed.
• Contract change at the TapTools boundary: configure() on conv_engine, stft, spectra::remapper and nr::reducer returns bool; load_ir() returns bool; a block size of 1 (N = 2, below the engine's minimum) is now refused where it used to run by accident.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QM4N2sNauJqVYyetTcgNcB

---

Generated by Claude Code