fix(metric-engine): make calibration serialization locale-independent - #85

Description

serializeCalibration() formats numbers with Number#toLocaleString() and no locale argument, so the thousands separator follows the host machine's default locale:

• en-US runner → dataset: 1,234 requests
• de-DE / de-AT (and many other locales) → dataset: 1 234 requests (a U+202F narrow no-break space)

This function emits a machine-embedded markdown comment (<!-- calibration: … -->) — a serialization artifact, not user-facing display. Locale-dependent output there is non-deterministic across developer/CI machines: it produces spurious diffs, can bake an invisible non-breaking space into committed files, and would break round-trip parsing of the block. It also makes the existing serializeCalibration unit test (which asserts dataset: 1,234 requests) fail on any non-en-US locale while passing on en-US CI.

Fix

Pin the serialization to en-US, the project's canonical format. en-US output is byte-identical (CI unaffected); all other locales become deterministic. User-facing toLocaleString() call sites (webview/UI) are intentionally left locale-adaptive.

-  lines.push(`  dataset: ${cal.datasetSize.toLocaleString()} ${cal.scope}`);
-  lines.push(`  flagged: ${cal.flaggedCount.toLocaleString()} (${cal.flaggedPct}%)`);
+  lines.push(`  dataset: ${cal.datasetSize.toLocaleString('en-US')} ${cal.scope}`);
+  lines.push(`  flagged: ${cal.flaggedCount.toLocaleString('en-US')} (${cal.flaggedPct}%)`);

Related Issues

None — found while running the test suite on a non-en-US (de-AT) machine, where serializeCalibration's test goes red.

Checklist

• npm run check passes (typecheck + lint + spellcheck + knip + tests)
• Changes are covered by tests (the existing serializeCalibration test now passes on every locale)
• Documentation updated (not applicable)

🤖 Generated with Claude Code