Summary

TestGenerateFromEvents drives the debounce logic through a real dockertest HTTP server and wall-clock time.Sleep calls, so the timing of debounce firings races the OS scheduler / timer resolution and the test fails nondeterministically (off-by-one or worse). It reproduces on main and is especially flaky on Windows. See #238.

This PR replaces it with TestNewDebounceChannel, which exercises newDebounceChannel directly inside a testing/synctest bubble. The fake clock makes the Min quiet-period coalescing and the Max cap fire at exact, reproducible virtual times, removing the real network and wall-clock dependence entirely.

Changes

• internal/generator/generator_test.go: replace TestGenerateFromEvents with TestNewDebounceChannel. Three deterministic subtests:
  • passthrough when Min == 0,
  • Min quiet-period coalescing (a burst collapses to one event, fired Min after the last event),
  • Max cap (output is forced out at Max while events keep arriving).

Notes

• No production code changes. newDebounceChannel already uses time.After, which synctest virtualizes.
• testing/synctest is stable as of Go 1.25 (go.mod declares go 1.25.5; the CI matrix tests Go 1.25 and 1.26). It is OS-independent, which removes the Windows timer-resolution flakiness at its root.
• Verified deterministic: go test -run TestNewDebounceChannel -count=200 passes.
• Trade-off: this focuses on the debounce timing (the logic that was actually flaky) rather than the full event → generate → output pipeline that the old test wrapped around a real Docker API server. If you'd prefer to keep an integration smoke test for that wiring (without timing assertions, so it stays deterministic), I'm happy to add one.

Fixes #238