fix(pi): preserve terminal and token accounting

[drewstone]

Problem

Pi sessions ending with stopReason: error were reported as successful, while reasoning and cache token dimensions were discarded.

The live GLM-5.2 run exposed both failures: two workers ended with Stream ended without finish_reason, yet trace analysis called both successful and omitted 30,397 root reasoning tokens plus 300,736 cache-read tokens across the root and active worker.

Change

• derive the root session status from Pi's final persisted message
• report terminal model errors as failed
• leave interrupted or unfinished sessions unknown instead of inventing success
• preserve reasoning, cache-read, and cache-write usage
• retain the native Pi stop reason on the root span

Proof

• pnpm test: 373/373 passed
• pnpm typecheck: passed
• pnpm build: passed
• replayed the three live run sessions: both errored workers now report failed; the externally interrupted leader reports unknown; root usage now includes 30,397 reasoning and 260,928 cache-read tokens