fix(verify): normalise the pytest summary duration when it carries a (H:MM:SS) suffix (AGT-4680)

src/verify/fingerprintStability.test.ts

@@ -67,6 +67,26 @@ describe('normalizeFailureOutput', () => {
     expect(a).not.toBe(normalizeFailureOutput(out('1.93').replace('1 error', '2 errors'), paths(SANDBOX_A)));
   });
 
+  it('normalizes a pytest summary that carries a (H:MM:SS) suffix (AGT-4680)', () => {
+    // pytest -q adds `(0:02:04)` once a run passes a minute, which every full cgf-portal
+    // suite does. The seconds were normalised but the suffix was not, so two runs of the
+    // SAME failure on main hashed differently and read as a new regression.
+    const out = (summary: string) => [
+      '=========================== short test summary info ============================',
+      'FAILED tests/test_b5_repository.py::test_the_empty_result_has_the_same_shape_as_a_real_one',
+      '!!!!!!!!!!!!!!!!!!!!!!!!!! stopping after 1 failures !!!!!!!!!!!!!!!!!!!!!!!!!!!',
+      summary,
+      '',
+    ].join('\n');
+    const a = normalizeFailureOutput(out('1 failed, 4234 passed, 581 skipped in 124.27s (0:02:04)'), paths(SANDBOX_A));
+    expect(a).toBe(normalizeFailureOutput(out('1 failed, 4234 passed, 581 skipped in 113.47s (0:01:53)'), paths(SANDBOX_B)));
+    expect(a).toContain('1 failed, 4234 passed, 581 skipped in <DURATION>');
+    // A different count is still a different failure.
+    expect(a).not.toBe(normalizeFailureOutput(
+      out('2 failed, 4233 passed, 581 skipped in 124.27s (0:02:04)'), paths(SANDBOX_A),
+    ));
+  });
+
   it('normalizes paths outside the command cwd — the subdirectory-cwd bug', () => {
     // A command with `cwd: packages/api` fails, but the traceback points at a
     // sibling package. Normalizing only the cwd left this line sandbox-specific.

src/verify/runner.ts

@@ -1,3 +1,3 @@
 import { execFile, spawn, spawnSync } from 'node:child_process';
 import { createHash, randomUUID } from 'node:crypto';
 import { constants } from 'node:fs';
@@ -233,8 +233,9 @@
     // carries no `=` decoration: `3 skipped, 1 error in 1.54s`. Left alone,
     // base and head fingerprints differed by timing alone, and every
     // pre-existing failure read as a new regression (vega-agent AGT-4118:
-    // the same ModuleNotFoundError on both sides, 1.93s vs 1.54s).
-    .replace(/^(\d+ [a-z]+(?:, \d+ [a-z]+)* in )\d+(?:\.\d+)?s$/gm, '$1<DURATION>')
+    // the same ModuleNotFoundError on both sides, 1.93s vs 1.54s). A run past a minute adds
+    // ` (0:02:04)` after the seconds, so that suffix is optional too (AGT-4680).
+    .replace(/^(\d+ [a-z]+(?:, \d+ [a-z]+)* in )\d+(?:\.\d+)?s(?: \(\d+:\d{2}:\d{2}\))?$/gm, '$1<DURATION>')
     .replace(/(Ran \d+ tests? in )\d+(?:\.\d+)?s/g, '$1<DURATION>')
     .replace(/(finished in )\d+(?:\.\d+)?s/gi, '$1<DURATION>')
     // pytest-xdist assigns the same failure to different workers on base and