fix(server): show new skills in the picker without restarting

[dakdevs]

Why

The $ picker keeps the first skill snapshot for a workspace until T3 restarts.

Evidence:

• refreshWorkspaceSnapshot returns the cached list once a cwd snapshot exists, so a later composer open does not scan again.
• Web and mobile only request that snapshot when the cwd is missing from workspaceSnapshots.
• The 5-minute health probe updates machine-level skills, but the composer reads workspaceSnapshots[cwd].skills first.

Fix: Stat each provider's listed skill roots every 3s. On a mismatch, drop that cwd snapshot (no publish) and reuse the one-shot refresh. Missing roots stay in the signature as path:! so creating a skills folder counts. fs.watch cannot do that.

Summary

New skills in a provider's skill directories show up in $ and / without restarting T3. Full discovery still runs only after a root signature moves.

What

• Each driver exposes the roots it actually loads via skillCatalogRoots.
• fingerprintSkillCatalogRoots stats those roots only. No BFS, no public force flag.
• A mismatch drops the cached workspace snapshot, then refreshWorkspaceSnapshot re-probes and publishes on the existing provider stream.

Who this helps

Anyone who adds a skill while T3 is already open: installing a skill mid-session, an agent writing SKILL.md, or copying a folder into .claude/skills / .agents/skills.

Test plan

• Open a thread so a workspace snapshot exists, add a skill folder under a provider root, wait a few seconds, open $, and confirm the new skill appears.
• Same path with /.
• Repeat on mobile New task and an existing thread.

[github-actions]

Thread transfer impact

⚠️ The latest CI run did not produce a thread transfer result for b537e62.

This comment will update automatically after the next completed run.