perf(cli): run outdated candidate searches in parallel

[harlan-zw]

Summary

• skilld outdated --system ran one HTTPS search per unmanaged Skill sequentially. 16 unmanaged Skills measured 10.1s wall; the same scan now measures 1.9s on up to 8 worker threads. WASI keeps the sequential path because it has no thread support.
• Compact output: no-match Skills collapse into one line and search failures group per message. The same scan went from 17 lines to 4.

Testing

• 177 workspace tests pass, clippy and fmt clean.
• New test proves the concurrency bound (max 8 in flight) and that 16 delayed searches finish far faster than the sequential time.
• Live binary timing before and after on this repository.

[harlan-zw]

Follow-up pushed in 5da2e82.

Scan width
outdated --system now also walks from the current directory up to the home directory, so Skills in a parent project directory are found. If the current directory is outside home, the walk stops at the filesystem root. Roots owned by a store scope that failed to read stay suppressed, so the ancestor walk cannot advise deleting copies it cannot verify.

Progressive output
outdated now reports each found Skill immediately (• name (agents, unmanaged)), rewrites a single ⠋ Checking name… spinner line while remote verification runs, and erases it before the final results. Non-terminal contexts and --json/--plain keep the silent path through a no-op reporter.