[cjfields]

Part of #31.

1. build_c.sh: C's timer on macOS (760e943)

• Problem: C's profiler (system/profiler_timer.c) reads the clock on macOS with three Mach calls, about 1.3 µs per alignment inside the timed interval.
  • C looked about 6× slower on 150 bp reads, and 20–25% slower on 1 kb fast modes (e.g. m1k_5 edit score: 113.5 → 91.5 ms).
  • So earlier M2 Rust/C ratios were too optimistic, including the old README ARM table.
• Fix: the macOS build now uses clock_gettime, as on Linux. The build fails rather than silently skipping the patch if the C source no longer matches.
• Testing: alignments are unchanged (differential suites core and biwfa: 284/284 identical).
• Cache stamp: cached C builds now carry a recipe number and are rebuilt when it changes, so older builds aren't reused. Every existing C build, Linux included, rebuilds once, in about a minute.

2. Performance docs (032646b, d1ed0f9)

• README: the March tables (Broadwell, Apple M-series, one dataset) are replaced by a summary of bench/sweep.py runs at fc51088 on an Apple M2 Pro, an AMD EPYC 7713 and an Intel Xeon Gold 6150. Each machine ran 301 cases, and output is identical to C everywhere. The summary also lists the known gaps:
  • 150 bp med/low;
  • Xeon affine2p CIGAR, cause identified in Performance: high-memory mode slower and larger than C on long pairs; port-quirk optimizations #31;
  • peak memory of high-memory CIGAR.
• docs/performance.md: the method, builds, datasets and full tables (by group, per algorithm, peak memory, overall), plus notes and how to reproduce. It's pinned to fc51088 and C 3282f83.
• bench/report.py: generates those tables from any set of sweep output directories.
• The README project tree and .claude/PLAN.md are updated for the new files.

The underlying data and the Xeon profile are posted in #31.

🤖 Generated with Claude Code

[cjfields]

These are largely documentation or documentation-related changes, so merging.