Partial deploy failure loses harness state — can create duplicate resources

## Problem

If deploying harness B fails after harness A succeeds, the partial state (containing harness A's new ID) is never persisted to disk.

**Flow:**
1. `HarnessDeployer.deploy()` returns `{ success: false, state: resultState }` — `resultState` contains harness A's ID
2. `actions.ts:531` extracts the state into `deployedHarnesses`
3. `actions.ts:541` returns early with an error — **before** `writeDeployedState()` at line 584

On next deploy attempt, the state file doesn't know harness A exists. The deployer takes the CREATE path (`createHarness` with `clientToken: randomUUID()`), producing a duplicate cloud resource. The previous harness A is orphaned.

## Expected behavior

Partial state should be persisted even on failure, so the next deploy attempt can UPDATE existing harnesses rather than re-creating them.

## Possible fix

Write deployed state before the early return at line 541:

```ts
if (harnessDeployError) {
// Persist partial state so successfully-created harnesses aren't orphaned
await configIO.writeDeployedState(buildDeployedState(...));
return { success: false, error: harnessDeployError };
}
```

## Context

Found during review of #1341.