fix(worktree): persist isolation origins for agent restore

[HaningZS]

Summary

Closes #298 by persisting the original repository alongside the actual isolated checkout, then resolving that durable recipe in the shared main-process spawn path.

• New isolated hires save worktreePath / worktreeOrigin in registry and roster, including voice/orchestrator hires.
• Team restore, Command Center restart/continue and sleep revival reuse a valid linked checkout, keeping dirty files intact.
• A missing checkout or a directory containing only .claude/ falls back to the saved origin, returns that corrected cwd to the renderer and clears stale isolation fields. The floor card says worktree gone — using base repo.
• Existing legacy linked checkouts can infer their origin. A dead legacy checkout with no saved origin fails explicitly with recovery guidance instead of guessing or provisioning into the dead path.
• A deliberate move to another cwd is not redirected by an old registry entry. A normal repository cannot be registered as a disposable linked worktree.

This chooses the issue's explicitly suggested origin-cwd fallback, not automatic recreation. It never prunes stale registrations, resets/deletes an agent/<id> branch, or removes a leftover directory. Existing unmerged commits remain available for manual recovery. #297's worktree exit policy and #299's CLI session-id capture remain separate concerns (the latter is in #665).

Before

Same 12 regression scenarios against upstream f44b986a production spawn and Restore Team functions: 1 pass / 11 fail. Images render actual node:test output; they are not GUI screenshots.

After

12 pass / 0 fail. Real Git subprocesses and temporary repositories with spaces, the real HiveManager and RosterStore, and the production main/renderer functions are exercised. The external CLI spawn and React mount are test boundaries, not real model invocations.

Validation

macOS / Node 24.13.0:

• 12 regressions: isolated hire, process restart with dirty files, missing checkout, .claude/-only residue, registry-independent roster recovery, missing legacy origin, live legacy origin inference, missing origin repo, intentional cwd relocation, repeated restore, normal-repo cleanup guard, renderer/roster round trip.
• Full suite with the independent catalog contract correction from test(models): validate hot and offline catalogs independently #669: 978 tests / 977 passed / 1 platform skip / 0 failed. A temporary test loader substitutes only test(models): validate hot and offline catalogs independently #669's catalog test; no production code or this PR's submitted files are changed by that integration check.
• Independent full-suite run: the only failure is the existing catalog equality assertion tracked by CI never runs the test suite; model-catalog-remote is failing on main #652/test(models): validate hot and offline catalogs independently #669, not a worktree test. This PR does not bundle that unrelated test fix.
• npm run typecheck, npm run build, git diff --check: passed.
• Final integration with test(models): validate hot and offline catalogs independently #669's catalog test correction and fix(memory): respect providers and quota cooldowns in reflection #671's independent reflector fix: 1004 tests / 1003 passed / 1 platform skip / 0 failed, typecheck/build passed, no merge conflicts. This is a real combined checkout, not a mocked production merge or substituted test file.
• Local socket tests require an execution environment allowing loopback/Unix sockets; the initial restricted run stalled on the existing hook-node test, then the complete run finished in ~43 seconds with those permissions.

Self-review

Provider-neutral path resolution, Node path helpers, and paths-with-spaces tests; no dependencies or permission escalation for agents. Dirty checkout contents, unmerged branch tips and stale Git registrations are explicitly asserted to survive fallback. The shared resolver is before provisioning, so an unknown/missing origin returns an actionable failure without rewriting registry/identity files. No claim of Windows/Linux GUI or real paid CLI session validation; those remain useful follow-up smoke tests.