feat(skills): track upstream for user-imported GitHub skills

[pjdoland]

Summary

User-imported GitHub skills can now opt into tracking upstream. The Import-from-GitHub dialog has a new Track upstream checkbox; ticking it stamps the installed skill with a tracks_upstream frontmatter flag and lights up a per-skill Sync button (and a panel-level Sync tracking skills batch action).

Sync re-fetches the bundle from the recorded source URL and replaces the on-disk content. Distinct from the existing managed-skills reconciler: tracking is opt-in per skill, never auto-scheduled, never auto-removed. The bundle on disk is preserved across every failure path (probe failure, tarball fetch failure, upstream repo deletion); the user clicked Sync, they're shown the error, and the existing bundle stays put.

Solution highlights

• Schema (skillset.py): two new optional frontmatter keys, tracks_upstream and tracking_ref. Default falsy and omitted from serialized SKILL.md when unset, so existing user-imported skills are unchanged on disk.
• SkillManager.sync_tracking_skill(scope, name): probes the commits API for the latest SHA at the recorded source, short-circuits when the SHA matches tracking_ref (no tarball fetch, no rmtree). When the SHA differs, stages the new bundle, atomically replaces, stamps the new ref. Probe failure raises rather than silently re-downloading. Tarball-fetch failure preserves the existing bundle on disk (rmtree only happens after staging succeeds).
• HTTP: POST /skills/{scope}/{name}/sync and POST /skills/sync-all-tracking route through run_in_executor. The batch endpoint runs per-skill syncs through an asyncio.Semaphore(3) so N skills don't take N * 15s in the worst case. Both gated by allow_github_skill_import (same network egress, same trust boundary). PUT /skills/{scope}/{name} accepts a tracks_upstream field for toggling on existing skills; that toggle is also gated by allow_github_skill_import so an admin's kill switch can't leak.
• Shared SHA helper (skill_github_import.py): new resolve_desired_sha(ref_info, token) consolidates the "if 40-char SHA short-circuit else probe" dance that the managed reconciler and the new sync path both run.
• Frontend: a "Tracking" badge (distinct from "Managed" by color and tooltip), a per-row sync button with disabled-state styling, a panel toolbar batch button, and an editor toggle for opting in/out post-import. The toggle flips optimistically and rolls back on PUT failure so the checkbox always reflects server state.

Mutual exclusion

A skill cannot be both managed and tracking. The manager refuses to enable tracking on a managed skill (ValueError → 400 at the HTTP layer). The two metadata pairs (managed_source/managed_ref vs tracks_upstream/tracking_ref) are independent on the wire but enforced exclusive at the manager.

Never-delete invariants

The "never delete" contract is enforced by the call ordering:

• rmtree(skill.root_path) runs only after stage_skill_from_github succeeds (which is the only step that can throw on a network or 404 error).
• A probe failure raises before rmtree is even reached.
• The reconciler is not extended to touch tracking skills; only the manual sync action can mutate them.

Testing

• tests/test_skill_manager.py: 12 new cases in TestTrackUpstream (import opt-in, toggle on/off via update_skill, sync happy path, sync skips on matching SHA, sync raises on probe failure, sync rejects non-tracking/managed/missing skills, sync preserves bundle on staging failure, list_tracking_skills filter) plus 3 cases in TestManagedFrontmatter for the schema fields.
• tests/test_skills_handlers.py: 9 new cases (SkillSyncHandler happy path, unchanged short-circuit, missing-skill 404, non-tracking rejection, 403 when imports disabled, 502 on probe failure, plus 4 PUT-toggle cases for managed/no-source/imports-disabled rejection and the regression pin that omitting tracks_upstream preserves the current value) plus 3 cases on SkillsSyncAllTrackingHandler.
• tests/ts/skill-wire-format.test.ts (new): 4 cases pinning the skillFromWire snake_case ↔ camelCase contract.

Verification: pytest tests/ --ignore=tests/test_claude_client.py → 905 passed. jlpm tsc --noEmit clean. jlpm jest → 185 passed. jlpm lint:check clean.

Risks and follow-ups

• Concurrent same-skill syncs (e.g. two browser tabs clicking the per-row Sync button simultaneously) race the rmtree/copytree. Sync-all is batched within a single handler so cross-skill is safe; only same-skill cross-handler is exposed. A per-skill lock on SkillManager is the principled fix; deferred to keep this PR focused.
• SkillEditor state sprawl: pre-existing trend (now 22+ hooks) amplified by this PR's toggle. useReducer would be the right refactor; deferred.
• No incremental progress feedback during sync-all: the UI shows "Syncing..." for the whole batch with no per-skill ticker. A future iteration could surface per-skill progress via streaming, but for the typical 1-5 tracking skills the current UX is fine.

Six-agent review pass (code-reuse, code-quality, efficiency, backend/contract, frontend/a11y, test architecture) ran on the four-commit base; the consolidated fix-before-merge items all landed in the final commit (5/5). Remaining items above are documented follow-ups with no immediate impact.