fix(watcher): stop ignored batches and small Flag::Rescan markers from breaking the index (#690)

[gustav-fff]

Closes #690

Root cause

Two independent bugs in handle_debounced_events (crates/fff-core/src/watcher/background_watcher.rs):

1. affected_paths_count summed raw debounced_event.event.paths.len() at background_watcher.rs:457 before ignore filtering. A burst of >1024 events under an ignored subtree (e.g. target/ churn from rustdoc / cargo build on macOS FSEvents recursive stream) forced a full rescan even though every path was filtered out at background_watcher.rs:451-453.
2. The paths.len() < 16 && all(!is_dir && !is_ignored) bypass at background_watcher.rs:365-374 silently breaked out of the event loop on Flag::Rescan without scheduling a rescan and without processing later events in the batch. On Linux, inotify Q_OVERFLOW arrives with no paths at all, so all() was vacuously true and every overflow marker was dropped on the floor. On macOS, any batch of <16 small non-ignored paths hit the same bypass.

Fix

• Count only actionable paths (those actually pushed to paths_to_remove / dirs_to_remove / paths_to_add_or_modify) against MAX_OVERFLOW_FILES. Ignored churn no longer trips the guard.
• Always schedule a full rescan on need_rescan(). Log event.info() alongside the paths so user dropped vs kernel dropped is distinguishable.

Architectural fix #2 from the issue (adaptive vs unconditional recursive FSEvents strategy) is intentionally out of scope — it needs benchmarking and a public strategy knob.

Steps to reproduce

Trigger 1 (ignored batch → false rescan):

```sh

Any recursively watched macOS repo with .gitignore containing /target

cargo doc --workspace # or any rustdoc generator producing many .js under target/doc

fff-search emits: "Too many affected paths in a single batch, triggering full rescan"

affected_paths_count = 1025, exactly matching the > 1024 branch,

even though every path is under /target and filtered out of the index.

```

Trigger 2 (silent Flag::Rescan bypass): unit-testable directly. On pre-fix main:

```sh
cd crates/fff-core

Feed a batch containing a single non-ignored, non-dir path with Flag::Rescan set.

Expected: watch_registry dispatches a Rescan event.

Actual: handler breaks out early; subscribers never see the rescan.

```

Both triggers are covered by unit tests added in this PR:

• ignored_batch_does_not_trigger_full_rescan — 2000 Create(File) events under .gitignored target/ must not dispatch a Rescan.
• rescan_flag_still_triggers_rescan_when_batch_is_small — a single-path Flag::Rescan must dispatch a Rescan.
• empty_path_rescan_flag_triggers_rescan — the Linux Q_OVERFLOW shape (empty path list) must dispatch a Rescan.

All three tests fail on origin/main (bug 1 dispatches a spurious rescan; bugs 2/3 silently return with no dispatch).

How verified

```sh
cargo test -p fff-search --lib background_watcher

5 passed; 0 failed — includes 3 new regression tests.

cargo test -p fff-search --lib

127 passed; 0 failed.

cargo test -p fff-search --test watch_subscription_test

11 passed; 0 failed.

cargo test -p fff-search --test new_directory_watcher_test

5 passed; 0 failed.

cargo fmt --package fff-search --check

clean

```

Automated triage via Gustav. Honk-Honk 🪿