Closes #744

Problem

v2 task IDs are UUID strings, but cf stats tokens --task and cf stats export --task typed the --task option as Optional[int]. Typer rejected any real v2 UUID at coercion time (is not a valid integer, exit 2), so per-task cost/token reporting and export were broken for every v2 task.

Fix

Type-only widening — no logic change:

• stats_commands.py: tokens() / export_data() --task → Optional[str]
• token_repository.py: get_task_token_summary / get_batch_token_usage → Union[int, str]
• metrics_tracker.py: get_task_token_summary passthrough → Union[int, str]

token_usage.task_id is already TEXT (schema_manager.py:172), so UUID rows round-trip unchanged; SQLite ?-binding never cared about the annotation. The break was purely Typer's int coercion.

Acceptance criteria (from #744)

• --task is str; get_task_token_summary/get_batch_token_usage accept Union[int, str]
• cf stats tokens --task <uuid> returns a saved UUID-keyed row

Tests

tests/cli/test_stats_task_uuid.py (new):

• cf stats tokens --task <uuid> → exit 0, correct totals (was exit 2 pre-fix)
• cf stats export --task <uuid> → writes the UUID-keyed record
• repository get_task_token_summary(<uuid>) → aggregated row

Verified live end-to-end: cf stats tokens --task <uuid> and cf stats export both render/export the saved UUID-keyed row. All 32 tests across changed modules pass; ruff clean.

Known limitations

• None. Legacy int task IDs still work (widened, not replaced).
• CLI docstring examples still show --task 1 (a valid str) — left as-is.