V3 shell: sync Tauri commands do blocking filesystem work on the main thread

[Loufe]

From the third rewrite code review (line references at rewrite 75c79c3).

Three commands are declared as sync fn, which Tauri 2 executes on the main thread, yet they do real IO:

• queue_add_paths (commands.rs:103) calls bridge.queue_add_paths, which runs crfty_engine::scan::expand_inputs (bridge.rs:440): a full breadth-first directory walk with an fs::metadata stat, path hash, and destructive_identity stat per file (scan.rs:43-135). Dropping a large folder tree freezes the UI for the whole walk.
• import_history (commands.rs:244) reads and parses a file (coordinator.rs:493-499).
• scrub_logs (commands.rs:252) rewrites every log file in place (logging/mod.rs:360).

The codebase demonstrably knows the rule and applies it selectively: open_path is async precisely so the desktop hand-off runs off the main thread (commands.rs:297-299), and check_for_update uses spawn_blocking (commands.rs:262-272). A native file open gets a worker thread while a recursive disk walk does not; that is inverted.

Secondary: every submit_* command blocks on a driver mpsc round-trip that includes journal append and fsync, and the driver does not service commands mid-compaction, so even cheap commands can stall the main thread behind disk work. Worth deciding deliberately how much of that is acceptable once the three commands above are fixed.

[Loufe]

Cross-reference: #51 adds a scrub_history shell command explicitly mirroring scrub_logs. Whatever threading shape fixes scrub_logs here (async command or spawn_blocking) is the pattern #51 should copy, not the current sync fn.

[Loufe]

#79 changes the History-import workload but not this bug:

• Import no longer stats paths, computes ContentKeys, probes media, or adopts verdicts.
• It still performs potentially large file read, strict JSON parse/validation, privacy transformation, reducer submission, journal append/fsync, and possibly conflict reporting.
• Those steps must remain off Tauri's main thread. Validate the document before mutation, then submit the native observations through the single writer.
• Import progress/summary terminology becomes inserted, duplicate, conflict, and rejected—not adopted/dropped.

#77 also adds History page queries. They should be async request/response commands and must not perform a full collection serialization on the main thread.