fix(worktree): never auto-discard a named agent's unintegrated work on PTY exit

[snehithareddy28]

What & why

Closes #297. When a named agent's terminal exits, teardownPty force-removes its isolated worktree — git worktree remove --force — with no check for work that was never integrated. The keep-if-unintegrated gate only existed for ephemeral workers, so the same teardown that preserves a worker's half-finished branch destroys a named agent's. As the issue says, this is the dangerous one of the restore family: the others break a restore, this one destroys work. The fix is the one the reporter asked for — apply the protection that already exists to the case it skipped.

One rule for every isolated agent. releaseWorktree() in git.ts removes a worktree only when it is clean and fully contained in the branch it was cut from; otherwise the worktree and its branch stay, and the caller is told why. It fails safe: whatever git cannot answer counts as work. Named agents now go through it on every teardown route — natural PTY exit, crash, kill from the UI, kill by voice. A preserved worktree is logged to log.jsonl, reported to god with the exact git worktree remove command to drop it by hand, and tracked in the same list the GC sweep already drains once the work lands in base. The agent's HIVE_ROOT/agents/<id> directory is its memory, not scratch, so the sweep never reclaims it alongside the worktree.

Two small enablers. The base branch is now recorded at spawn (worktreeBases) instead of computed and thrown away; a worktree tracked without one is compared against the parent repo's current branch, and kept when even that is unknown. And the GC sweep skips an id that is live again as a named agent (restore re-enters the worktree by cwd), not only a live worker.

Type of change

• Bug fix
• New feature
• Refactor / cleanup
• Docs
• Build / CI

Evidence

Before

After

---

Notes for review:

• Five tests in a new test/worktree-release.test.cjs, each on a real repo with a real agent/<id> worktree: clean and integrated → removed (and gone from git worktree list); uncommitted changes → preserved with the file still there; commits the base lacks → preserved, branch intact; the same worktree released once god merges the branch; an unanswerable base → kept.
• finalizeWorkerWorktree is untouched; workers keep their scratch-dir handling. finalizeAgentWorktree is its named-agent sibling with scratchDir: null, and the two removeWorkerScratch calls in the sweep are gated on that field.
• On the rest of the restore family (Isolated pi agent dir has no models.json, so the first session dies ~415ms after every hire #296, Worktree origin is never persisted, so restore falls back to the dead worktree path and agents stay "restorable" forever #298, pi session_id is never forwarded by hive-bridge.js, so a restored pi agent always starts a fresh conversation #299): this PR only stops the destruction. After an app restart the in-memory worktree maps are empty, so a restored agent's worktree is never torn down by this path at all — Worktree origin is never persisted, so restore falls back to the dead worktree path and agents stay "restorable" forever #298 (persist the origin) is the piece that would let restore find the worktree again; it is unchanged here.
• Behaviour change to be aware of: killing a named agent from the UI no longer removes a worktree with unintegrated work. It is reported to god, listed with the preserved worker worktrees in the Workers tab, and reclaimed automatically once its branch is merged.
• npm run typecheck and the full npm run test:focused suite (839/839 on this branch) pass locally.

Discord: asr2805