Add r8brain-free-src to the resampler comparison

[tap]

What this changes

Adds r8brain-free-src (7.5, pinned at commit 9e73d2dd in cmake/r8brain.cmake) to every part of docs/COMPARISON.md: the quality notebook, the host throughput bench and the embedded instruction counts. The embedded counts now report steady-state cost separately from one-time construction.

Why

We had no measured comparison against r8brain, a widely used MIT resampler. Re-measuring also turned up stale numbers in the comparison. The old embedded per-frame figures included one-time construction, so the headline "~9.8× cheaper than libsamplerate on M33" no longer matched the library.

Verification

All of this was built and run in one container on 2026-09-25/26:

• Quality: notebooks/asrc_comparison.ipynb is re-executed and committed executed, and its assertions pass.
  • r8brain CDSPResampler24 measures −143.9 dB THD+N at 24-bit I/O, −150.8 dB at float I/O, and 149.1 dB DR.
  • The new "Latency vs. passband" cell sweeps r8brain's transition band. At the default it holds back 789 input frames. The lowest-latency setting still flat to 20 kHz is 8 % at 200 frames, and the ~1 ms settings are −35 dB at 20 kHz.
• Host throughput: srt_bench_compare, median of 5, all engines in one session. The numbers are in COMPARISON.md.
• Embedded counts: QEMU plus the counting plugin, all 36 runs with ok=1, on M55 (mps3-an547), M33 (mps2-an505) and Hexagon (qemu-hexagon built with plugins from the checksum-pinned 8.2.2 source, toolchain 19.1.5). The toolchain and plugin-header pins match compare.yml.
  • The libsamplerate totals under the old metric reproduce the previous table exactly: 2,218 / 6,400 on M55, 49,424 / 149,426 on M33, 9,102 / 26,959 on Hexagon.
  • I have not dispatched compare.yml on GitHub. The recorded counts come from this local reproduction, and one manual run would confirm them on CI.
• Default build: configured with -DSRT_WERROR=ON; all 73 tests pass.
• compare-smoke steps: run locally in fresh directories, both pass.
• Formatting and tidy: clang-format 18.1.3 is clean on all changed C++. clang-tidy is clean on the new shim.
• clang -Werror: my new code builds clean. The build still fails on existing lines in bench/bench_asrc.cpp:83 and in the old srtBench in bench_compare.cpp, both int64_t/size_t signedness. I didn't touch these, and CI doesn't build them with clang.

Notes for the reviewer

• Notebooks re-executed. SampleRateTap's own row moved from −132.1 dB to −133.9 dB. The notebook hadn't been re-run since the compensated prototype landed (a45043a, 2026-07-04), so this is a re-measurement, not a change this PR makes.
• Metric change in the embedded table. Each engine now builds at 2 s and 4 s. The difference is the steady-state cost per frame and the remainder is construction.
  • Steady state, balanced float beats r8brain 1.1–2.2× on every embedded target.
  • Q15 on M33 is 879 insn/frame, against ~26.6k for r8brain and ~49.2k for libsamplerate MEDIUM.
  • Construction is where we lose: about 1.3 G instructions on M33 for the soft-double filter design, against tens of millions for the competitors. The doc says so plainly, and it may deserve its own issue.
• r8brain on x86. r8brain is faster than balanced on x86 (1.1–1.7×), at 8–33× the filter delay. The doc states this too.
• No-op mutex for Cortex-M. The Cortex-M comparison builds force-include bench/icount/r8b_single_thread_mutex.h. r8brain's filter cache uses std::mutex with no hook to replace it, and thread-less newlib doesn't declare one. The stand-in is exact for the single-threaded workload and off the per-sample path.
• Fetch method. r8brain is fetched with GIT_REPOSITORY plus a full commit SHA rather than a tarball URL_HASH. Its newest tag (6.5) predates 7.x, and the archive download was blocked in my environment. It's comparison-only and never linked into the library; the README provenance section now says so.
• Stale text left for a follow-up. Only the "9.8×" sentence in book/src/part0/two-crystals.md is corrected here. Other stale figures (−132.1 dB, 5,043 insn/frame, June host throughput) remain in the book (two-crystals.md, budgets.md, cortex-m.md, notebooks.md, hardware.md, glossary) and in the Pico example READMEs.

🤖 Generated with Claude Code

https://claude.ai/code/session_01G3HxzEGiZp7jStoMuYNuBK

---

Generated by Claude Code