[NaNIsaac]

Corrects the cause I gave in the issue text of #42 and fixes the flake in the test. Tier: low (tests only; nothing in src or rust/src changes).

Fixes #42.

What the issue got wrong: it said the Rust build truncates the command time to whole milliseconds while the Python build rounds it. That is true only of Python's shell-hook ingestion (commands/term.py, which reads $EPOCHREALTIME). cdg cap, which the test runs, measures in core/runner.py with int((time.monotonic() - started) * 1000), and Rust with started.elapsed().as_millis(): both truncate, so the two builds agree on how a given elapsed time is stored.

The flake is still real and has a different source: the test runs the command once with each build, and each run is timed separately. A command that takes about a millisecond (sh -c 'echo ...; exit 1' on a fast runner) can come out as 0 ms in one run and 1 ms in the other, and an entry with 0 ms has no duration attribute. Making Rust round would not remove this, only move the threshold to 0.5 ms.

• normalise in rust/tests/capture_common/mod.rs now drops the duration="..." attribute instead of masking its value, so the comparison no longer depends on whether a run was fast enough to have one. stable() in the same test file already removes duration_ms from the stored JSON for the same reason.
• New test a_command_that_took_under_a_millisecond_compares_equal_to_one_that_took_longer: the same <command> with no duration, 1ms and 1.5s normalise to one string.
• Rendering of a stored duration is still covered by the unit test in rust/src/term.rs (duration="1.5s").

Tested locally on Windows: cargo fmt --check, cargo clippy --all-targets -- -D warnings, cargo test pass; cargo test --test cap -- --include-ignored with the Python build from uv sync --python 3.15 and CODIGEST_PARITY=1: 15 passed. The failure itself needs a Linux runner and a sub-millisecond run, so it was not reproduced here; the new test checks the comparison directly instead. The ubuntu parity job is the real check.

[NaNIsaac]

local-ci: b74b5e1 | ruff n/a (no Python change) | cargo fmt, clippy -D warnings and cargo test ok (cap parity with the Python build: 15 passed, Windows) | readme n/a