Add block-size study notebook (32/64/240) with latency comparison

[tap]

Companion to asrc_demo.ipynb: a full study of processing block size (32 / 64 / 240 frames = 5 ms) against latency and quality, committed with executed outputs (5 figures), every claim assertion-pinned.

What it shows

• Lock & breathing: all three sizes lock and run 30 s at +200 ppm with zero underruns; steady-state FIFO fill oscillates by almost exactly one block (0.60 / 1.28 / 4.99 ms peak-to-peak) — latency breathing, the Track stage's designed response to block quantization.
• Quality, with a calibrated instrument: a phase-noise-style decomposition splits the residual at ~60 Hz into pitch modulation (cents) and a wideband floor. The instrument is self-tested in-notebook on synthetic FM and recovers 1.000 cents / 111 dB exactly. Measured: ~0.87 cents rms FM over a 60.9 dB floor at B=32 → 1.34 cents / 53.1 dB at B=240, with B=240's FM peaks (4.2 cents) below but within sight of the ~5–8 cent audibility region.
• Latency, empirically: per-impulse transit times through the live converter show measured latency as a band — mean ≈ designed − B/2 (the pull-instant backlog runs half a block under the setpoint), width ≈ the breathing. A stacked decomposition makes the economics plain: 0.5 ms fixed filter delay vs block-driven setpoint (1.83 / 3.17 / 10.50 ms designed). Block size is the latency budget; the ASRC is a rounding error.

Honesty notes (the interesting part)

• Getting the quality measurement right took four iterations, each caught by the in-notebook calibration cell: boxcar passband droop leaked 1.6% of the modulation into the noise figure (bounding it at 63.5 dB), and signal-domain reconstruct-and-subtract fails structurally because sub-split phase errors multiply the carrier. The shipped instrument works in the demodulated domain with a flat-passband FFT split. The failed approaches are documented in the cell's docstring so nobody retraces them.
• The measurements showed the README's "sub-cent wow at sub-hertz rates" wording undersold coarse-block behavior — corrected to the measured figures, with a link to this notebook as the evidence.
• The summary is blunt about regimes: coarse-block operation is serviceable (monitoring/comms; vastly better than naive-FIFO clicks at ~30 dB) but not the transparent ~126 dB regime of fine-grained transfer; the guidance is to push in sub-blocks when transparency matters.

https://claude.ai/code/session_01HuAFfoeD5a5Xe5aGNA16M9

---

Generated by Claude Code