[dajiaohuang]

Summary

• Keep the original latency timer when additional requests start while the loading bar is already active.
• Add fake-time regression coverage for concurrent starts crossing the initial threshold.

Previously, each start() emitted a state that caused switchMap to cancel and restart the timer. A sustained sequence of overlapping requests could therefore indefinitely postpone the bar even after the first request exceeded latencyThreshold.

Validation

• Core tests: 2 suites, 20 tests passed (jest packages/core/src --runInBand --coverage=false).
• Prettier check passed for both changed files.
• git diff --check passed.

Tests use synthetic progress and fake timers only; no application, browser profile, or external service was accessed.