test: leave the duration attribute out of the build-to-build comparison

rust/tests/cap.rs

@@ -15,6 +15,21 @@ fn history(home: &Path) -> Vec<Value> {
         .collect()
 }
 
+#[test]
+fn a_command_that_took_under_a_millisecond_compares_equal_to_one_that_took_longer() {
+    let quick = r#"<command index="1" exit_code="1" at="2026-10-17T01:58:14+09:00">"#;
+    let slower =
+        r#"<command index="1" exit_code="1" duration="1ms" at="2026-10-17T02:00:00+09:00">"#;
+    let longer =
+        r#"<command index="1" exit_code="1" duration="1.5s" at="2026-10-17T02:00:00+09:00">"#;
+    assert_eq!(normalise(quick), normalise(slower));
+    assert_eq!(normalise(quick), normalise(longer));
+    assert_eq!(
+        normalise(quick),
+        r#"<command index="1" exit_code="1" at="<TIME>">"#
+    );
+}
+
 #[test]
 fn output_from_both_streams_is_shown_and_stored_and_the_exit_code_passes_through() {
     let home = tempfile::tempdir().unwrap();

rust/tests/capture_common/mod.rs

@@ -77,11 +77,13 @@ pub fn require_parity() {
 }
 
 /// What a command printed, with the parts that legitimately differ between two runs
-/// replaced: timestamps, durations, relative times.
+/// replaced: timestamps, durations, relative times. A `duration` attribute is dropped, not
+/// masked, because both builds store whole milliseconds and omit the attribute at 0, so two
+/// runs of a command that takes about a millisecond can differ in whether it is there at all.
 pub fn normalise(text: &str) -> String {
     let patterns = [
         (r"\d{4}-\d\d-\d\dT\d\d:\d\d:\d\d[+-]\d\d:\d\d", "<TIME>"),
-        (r#"duration="[^"]*""#, r#"duration="<D>""#),
+        (r#" duration="[^"]*""#, ""),
         (r"\(\d+ \w+ ago\)", "(<AGO>)"),
     ];
     let mut out = text.replace("\r\n", "\n");