[bdbarnett]

Fixes #290.

#284 let a splice search out to a whole window when nothing within half a window matched the other tap. It also did that where the other tap read only the last few LSB of a fading repeat, or silence. There, rounding alone makes every match look poor, so the search went the whole way out and left the starting tap a window off the material that came next. Through a pitch shifter with feedback, the third pass of a short burst came back 19 dB under what Feedback leaves it.

Now the far search also needs the other tap to read more than 8 LSB on average over the compared stretch. A held chord reads far above that, so everything #284 fixed stays fixed. Nothing changes in the API.

Regression, not a corrected artefact

audioeffects' PitchShifter at 22,050 Hz, -5 semitones, a 30 ms window and Feedback 0.3, pass 3 against Feedback² of pass 1:

	before #284 (9321f25)	#284 (c03a357)	this branch
pass 2	+1.2 dB	+1.2 dB	+1.2 dB
pass 3	+2.5 dB	-19.0 dB	+2.5 dB
PitchShifter's repeats check, all 560 cells and seeds	0 fail, worst -8.3 dB	16 fail, worst -19.0 dB	0 fail, worst -5.3 dB

The old reading was not inflated: each trip gains the same 1.2 dB, so +2.5 on pass 3 is two trips of it. Under #284 the energy of pass 3 had gaps in it. A trace of the splices showed why. When the far search fired, the other tap read 42 to 84 in total over 48 frames (about 1 to 2 LSB each), and once it read 0. Each time it moved the tap 660 frames.

The check also passes with the burst 20 and 40 dB quieter, where #284 still failed at -20 dB.

#246 stays fixed

	before #284	#284	this branch
E19, worst note of the seven cases	-60.2 dB	-1.4 dB	-1.4 dB
E18 A2 and A#2	-5.2, -5.7 dB	-5.3, -5.3 dB	-5.3, -5.3 dB
E18 C3 triad	-2.9 dB	-0.8 dB	-0.8 dB
450 random chords (+12 at 48 kHz, +5 at 44.1, -7 at 32), worst note: median / 10th pct / more than 10 dB down / worst	-4.5 / -14.9 / 117 / -56.2	-1.4 / -11.7 / 61 / -41.9	-1.4 / -11.7 / 61 / -41.9

Every E19 note reads the same to 0.01 dB as under #284, and so does the 450-chord sweep. The options rows that #284 moved (audioecho.FeedbackDelay@options, with and without damping) render byte for byte as they do under #284.

Tests

• New trait E20 in tests/test_cpython_audioecho.py: a 40 ms burst through audiodelays.PitchShift and then the loop shifter, with feedback, as a pitch shifter wires them, at 14 onsets across one splice period. Each repeat must be within 6 dB of Feedback² of the one before. It fails on c03a357 (-17.6 and -20.0 dB at two onsets) and passes here (-3.7 dB at worst).
• tests/test_cpython_audioecho.py, test_cpython_free_threading, test_binding_parity, test_cpython_audiomodal and test_cpython_audioshaper pass (117 tests).
• verify_dsp.py, CPython against a standalone MicroPython v1.29.0 unix build from this branch: 58 comparisons, 0 failures. The E20 chain renders byte for byte the same on both at all 14 onsets.

What's left

CircuitPython wasn't built for this change, so the three-way run is still to do. The change is one integer comparison, made before the search widens.