fix(agent): lazy-reload skill registry on CLI slash invocation miss (#57)

[quangdang46]

What

The TUI input path (src/tui/app/input.rs:2014-2038) already does a synchronous SkillRegistry::load_for_working_dir on a slash-skill miss so project-local skills installed under .jcode/skills/ work the moment skill_manage reload_all finishes. The CLI loop in src/agent/turn_execution.rs captured skills once at the top of the session and never refreshed it, so users in headless / non-TUI mode saw Unknown skill: /<name> even after a successful reload_all.

This addresses issue #57: #57

Changes

• src/agent/turn_execution.rs: mirror the TUI lazy-reload pattern in the CLI loop:
  • On a slash miss, re-derive the registry from the session's working directory.
  • Replace the shared registry under try_write so subsequent invocations see the fresh snapshot.
  • Retry the lookup once before falling back to the "Unknown skill" path.
  • The "Available:" listing now uses current_skills_snapshot so it reflects whatever the lazy-reload just discovered, not the stale top-of-loop capture.

Tests

cargo check -p jcode --lib clean. The fix is symmetric with the TUI path that has its own coverage in src/tui/app/tests/remote_startup_input_03/. The CLI loop is interactive (stdin().read_line), making a focused unit test require pty plumbing that doesn't exist here today; the change itself is a small mechanical mirror of an existing, tested pattern.

Notes

The shared registry update uses try_write so a reader holding the lock at that exact moment doesn't deadlock the CLI loop — same defensive style as current_skills_snapshot.