Brandon ThomasandClaude Opus 5.5 715f85c612 Tests: linear-time checks measure CPU time, not wall clock under load (#850) ...

The five tests that assert linear layout/export/hyphenation by comparing
the cost of an n-sized and a larger input (layout: unequal formats, hidden
separators with auto hyphenation, hidden float anchors; pdf: fragmented
hidden headings; proof: hyphenating a long word) timed both with wall
clock, fastest of three. On a loaded machine (CI, shared dev boxes, the
test harness running other tests in parallel) a preempted big run, timer
resolution on the small run, or a tight 4x/8.0 limit (hyphenation measured
~7.2 when linear) made them fail now and then.

Each now uses a shared cost_ratio helper (test-only, no production change):
- CPU time of the test thread from /proc/thread-self/schedstat on Linux, so
  time spent waiting for a CPU is left out (wall clock elsewhere);
- the small input runs `scale` times per timing so both timings are long
  against the clock's resolution;
- small and big runs alternate, five rounds after a warm-up, cheapest kept;
- 8x the input everywhere: linear ~8x, quadratic ~64x, limit 24x.

Measured under load (load avg ~15-35 on 32 cores): 60 linear ratios in
7.7-8.8. With the old quadratic code injected locally each test fails:
unequal formats (uncapped style bucket scan) ~90-104x, hidden separators
(hyphenator rescan per point) ~36x, float anchors (run walk restarted per
anchor) ~188x, PDF headings (left-out range scan per char) ~37x,
hyphenation 50-65x.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

2026-10-12 04:37:03 -04:00

..

src

Tests: linear-time checks measure CPU time, not wall clock under load (#850)

2026-10-12 04:37:03 -04:00

Cargo.toml

M15: Polish interface, and Polish spelling, grammar and hyphenation (#346)

2026-10-12 04:04:03 -04:00