#1098 Report the true watch mode when the events directory is missing

[williamthorsen]

What

Fixes an issue on Linux where Fleet's startup line claimed an active watch on its events directory even when that directory did not exist. The line now reports when Fleet is falling back to periodic rescans instead, naming the directory it could not watch and the reason.

Why

The startup line is the operator's only signal for whether low-latency updates are in effect, and the module's stated contract is that a fallback to periodic rescans is always announced. On Linux that promise was not being kept, and it was the watcher suite's failure on CI that surfaced it.

Details

🐛 Bug fixes

• The watch mode is decided from a direct probe of the target directory rather than from whether starting the watch reported an error, so a missing directory produces the fallback line on every platform.
• The fallback line names the directory and the underlying reason, replacing wording that attributed every failure to the watch API even when the directory was simply absent.

♻️ Refactoring

• startWatcher resolves its watch through an injectable starter, matching the seams already used by createGitAdapter and createGithubAdapter. Production omits it and takes the default.

🧪 Tests

• The debounce path and the mid-run watch-error downgrade are covered, both driven through an injected starter rather than real filesystem events.
• A test asserts that a missing directory is detected before any watch is attempted, pinning the probe ahead of the watch.
• One test exercises the default starter against an existing directory, so the production path stays covered.
• Temp directories and watchers are scoped to the test that creates them.

Closes #1098

[github-actions]

Dependency audit

Production dependency audit passed.