[ucgmsim]

Closes #237

What was wrong

significant_duration::significant_duration expects a cumulative Arias
intensity array (non-decreasing along each row, total in the last
element) — threshold_search binary-searches it. The Python binding
(_significant_duration in src-rust/lib.rs) always computes
cumulative_arias_intensity first.

bench_significant_duration in benches/intensity_benchmarks.rs instead
passed the raw sinusoidal output of generate_waveforms straight into
significant_duration. Those rows go negative and are not monotonic, so:

• for stations whose last sample is negative or below 1e-10,
  threshold_search took its early 0.0 return and did no work;
• for the rest, the binary search ran over unsorted data.

The benchmark therefore measured neither the real workload nor a
meaningful path, and left out the cumulative-integration cost that
dominates _significant_duration in practice.

The fix

Extracted the waveform generator and a new
significant_duration_from_waveforms helper (integrate, then search — the
same two steps the Python binding runs) into benches/common.rs, shared
between the benchmark and a new test. bench_significant_duration now
calls the helper inside b.iter, so it measures the full Ds pipeline as
Python uses it, and the benchmark group is renamed to
Significant_Duration_Full_Pipeline to say so.

How the new test shows it

tests/significant_duration_benchmark.rs calls the shared helper on
synthetic waveforms (same generator as the benchmark) and asserts every
station's Ds duration is in (0, total_duration]. Before the fix (i.e.
with the helper calling significant_duration directly on raw waveforms,
as the benchmark used to), this test fails — reproducing exactly the
0.0-duration degeneracy the issue describes:

station 3: expected a Ds duration in (0, 9.995...], got 0; feeding raw
waveforms directly into significant_duration (instead of their cumulative
Arias intensity) collapses this to 0.0 for most stations

After computing cumulative_arias_intensity first, it passes.

What to check

• benches/common.rs is a new file included into both the bench binary
  (#[path = "common.rs"] mod common;) and the test
  (#[path = "../benches/common.rs"] mod common;) — this is the only way
  to get a real (non-smoke) pass/fail check on benchmark-only code, since
  cargo test --benches just runs criterion's panic-only smoke pass and
  doesn't check output values.
• No production (src-rust) code changed; this is benchmark/test-only.

Cost $1.59, 15 turns.

🤖 Generated with Claude Code