Correct the M33 budget (stereo Q15 fits one core); refresh stale figures; fix clang sign-conversion

[tap]

What this changes

Follow-ups to #45:

• M33 budget correction. The README, book, HARDWARE_TESTING.md and both Pico examples now state per-frame M33 costs as steady state (servo and FIFO included): 1,138 insn/frame for stereo Q15 and 3,326 for 12 channels. They used to say 5,043 and 10,027. That changes a conclusion: 48 kHz Q15 stereo fits one 150 MHz core; the old text called it infeasible.
• Stale figures refreshed. The book's −132.1 dB (now −133.9) and its June host-throughput numbers now match docs/COMPARISON.md.
• Construction cost qualified. Where the text prices construction in seconds, it now notes that the RP2350's DCP coprocessor makes a Pico 2 cheaper than the soft-double count suggests. This is already recorded in docs/PERFORMANCE.md.
• SRT_SC_SECONDS (default 2) lets the ratchet workload be built at 4 s, so steady state can be separated from construction.
• clang -Werror fix. Two benchmark SetItemsProcessed calls had a sign conversion that failed clang -Werror.

Why

The old figures divided a 2 s pipeline baseline by its 96,000 frames. That quotient also carries the converter's one-time construction, which is soft-double filter design on the M33, so it overstated the per-frame cost about 4×. I built the June commit c609a0f the same way: its per-frame costs are the same 1,138 / 3,326. The old numbers were inflated from the start. Only construction has grown since (375 M → 1.31 G with the compensated design).

Verification

• Steady-state counts. M33 counts come from QEMU with the counting plugin (same plugin header SHA as CI).
  • Today: pipeline_q15 1,417,933,654 at 2 s / 1,527,220,570 at 4 s; pipeline12_q15 1,896,402,851 / 2,215,675,798.
  • June c609a0f: 484,257,006 / 593,544,599 and 962,728,307 / 1,281,999,472.
  • My 2 s counts sit +0.36 % above the committed baselines, a constant ~5.05 M on both workloads. It looks like a startup-code difference between toolchains, and it cancels in the difference.
• Ratchet binaries unchanged. The gated binary's .text is byte-identical to main, both at the default and at an explicit SRT_SC_SECONDS=2, so no baseline moves.
• Book build. It builds with CI's pinned mdBook v0.4.40 (checksum verified) with 0 warnings/errors, so no stale anchors.
• clang. A clang -Werror build of every optional target (benchmarks, compare bench, C ABI, shim, icount compare) now has 0 errors.
• Not re-run. The test suite, since no library code changed. The Pico firmwares are unbuilt (no Pico SDK here); only their comments changed, and clang-format is clean.
• Not verified on silicon. The budget statements are still instruction-space, pending the pico2_cyccnt measurement.

Notes for the reviewer

• Contract change (documentation): the README's M33 deployment guidance flips from "stereo wants fast() or the second core" to "stereo fits one core". The 12-channel conclusion (16 kHz in the dual-core example) stands, on a much smaller margin: 3,326 against 3,125 at one instruction per cycle.
• Pico calibration. pico2_cyccnt's calibration now divides cycles/frame by the steady-state counts. The old denominators included setup the firmware never times.
• Related DspTap change: tap/DspTap has a bit-identical change that cuts the M33 filter design by 32 % (~400 M instructions per construction). Bumping this repo's pin to it will lower every construction-inclusive icount baseline beyond ±3 %, so that bump needs re-recorded baselines from CI logs.

🤖 Generated with Claude Code

https://claude.ai/code/session_01G3HxzEGiZp7jStoMuYNuBK

---

Generated by Claude Code