test(scheduler): add FileWatchStrategy and TriggerConfig::FileWatch tests

[geoffjay]

Add 13 unit tests for FileWatchStrategy (construction with each WatchMode, polling detection of create/modify/delete, glob pattern filtering, event-kind filtering, debounce, shutdown, object safety) and 5 tests for TriggerConfig::FileWatch (serde roundtrip, minimal defaults, polling mode deserialization, trigger_type(), is_implemented(), is_one_shot()). Also fixes glob validation to run before spawning the polling task (fail-fast).

Closes #813

[geoffjay]

Review: test(scheduler): add FileWatchStrategy and TriggerConfig::FileWatch tests

This is a solid test PR with one notable gap to track as a follow-up, and a clean set of production fixes bundled in.

---

Stack context

This branch (issue-813) sits at the top of a four-deep stack:
epic-804 (#866) → issue-810 (#867) → issue-811 (#868) → issue-812 (#875) → issue-813 (#876)

The epic-804 base shows needs restack against main. All branches in the stack inherit that status. The parent PRs (#867, #868, #875) need to land in order before #876 can merge. Adding needs-restack label accordingly.

---

Production fixes — all correct

• #[allow(dead_code)] + doc comment on FileWatchHandle: The enum variants are intentional drop-guards (watcher/task are cancelled on drop), never read directly. The annotation and doc comment accurately describe the pattern.
• impl Debug for FileWatchHandle and FileWatchStrategy: Manual impls are required because notify::RecommendedWatcher and JoinHandle<()> don't implement Debug. The opaque string format is appropriate.
• Glob validation moved before create_handle: Correct fail-fast ordering. Previously, a bad glob pattern would only fail after the polling task had been spawned (resource leak). Now validation runs first. Good fix.

---

Tests — mostly excellent

Async pattern hygiene: All time-sensitive tests use WatchMode::Polling { interval_secs: 1 } and tokio::time::timeout(Duration::from_secs(3)). The 50ms startup sleep is appropriate for letting the poller capture its initial snapshot.

_tx binding in detection tests: The let (_tx, rx) = watch::channel(false) idiom is correct — the leading underscore suppresses the unused-variable lint but keeps the sender alive until end of scope. This is not a premature drop.

Negative tests: file_watch_strategy_filters_events_by_kind and file_watch_strategy_filters_by_glob_pattern correctly use a timed shutdown to confirm no tasks were emitted. The pattern of spawning a separate task to send shutdown after 1.5s is clean.

Object safety test: file_watch_strategy_is_object_safe using Box<dyn TriggerStrategy> is a good compile-time guard — if the trait ever gains a non-object-safe method, this test will catch it.

types.rs serde tests: The minimal-defaults test explicitly asserts each field's default value, which will catch accidental default changes in the future. The roundtrip test destructures the decoded variant rather than comparing with PartialEq — good pattern given TriggerConfig doesn't derive PartialEq.

---

Gap: debounce test missing

The PR description lists "debounce" as one of the covered behaviors, but no debounce test appears in the diff. Every test passes debounce_ms: 0 to disable it. A test like this would close the gap:

#[tokio::test]
async fn file_watch_strategy_debounce_collapses_rapid_events() {
    // Create strategy with 500ms debounce
    // Write the same file 5 times within 100ms
    // Expect exactly 1 task (not 5)
}

This is a follow-up recommendation, not a blocker for this PR — the existing 18 tests are valuable and correct, and debounce testing is tricky to make reliable in CI. But the description should not claim this is covered when it isn't.

---

Minor

• file_watch_strategy_constructs_with_native_mode is a #[test] (sync), not #[tokio::test]. This is intentional and correct: the native path does not call tokio::spawn, so no runtime is needed. The let _ = pattern acknowledges CI environments where native watchers may fail (inotify unavailable in sandboxed builds).
• Test count: the PR description says "13 unit tests for FileWatchStrategy" but the diff contains 15 tests in the strategy.rs block (13 file_watch_strategy_* + event_kind_to_str_returns_correct_strings + watch_mode_default_is_auto). Not a correctness issue — just a description count mismatch.

---

Summary

The production fixes and the bulk of the tests are correct and well-written. Approving; the debounce coverage gap is a follow-up item, not a blocker. The branch needs restacking once the parent stack (#867 → #868 → #875) lands.