fix(pi): run memory learning in background

[OLDyade]

Summary

• enqueue Pi turn-end extraction, skill evolution, Dream, and learned-skill sync instead of awaiting them inside the foreground agent_end hook
• preserve strict FIFO ordering and the originating session/model context for every background turn
• report background failures through Pi native UI notifications and reject session shutdown after the final sweep
• drain pending work before session_shutdown completes so normal exits do not lose memory writes

Why

Pi waits for extension agent_end promises before clearing its foreground working state. A real local run completed the assistant response at 16:49:07 but remained working until MemFlywheel finished at 16:50:15, adding about 68 seconds of visible latency.

Validation

• pnpm run test — 261 tests passed
• pnpm run format:check
• git diff --check
• real Pi 0.83.0 E2E using the configured deepseek/deepseek-v4-flash model:
  • assistant response completed at 17:00:00.230
  • Pi emitted agent_end and agent_settled without waiting for memory work
  • typed memory was written asynchronously at 17:00:03.979
  • Dream/index work continued in the background without holding the foreground lifecycle
  • a fresh session used Pi native read on the typed memory body and recalled the exact marker

Lifecycle contract

agent_end -> snapshot + enqueue -> return immediately
background queue -> extraction -> skill evolution -> Dream -> skill sync
session_shutdown -> drain queue -> final sweep -> cleanup

The foreground is eventually consistent with the memory store: a new session started immediately after an answer may begin before that turn has been extracted, while same-session conversation context remains available.

[FenjuFu]

LGTM. Moving Pi's turn-end work (extraction / skill evolution / dream / learned-skill sync) off the foreground agent_end hook into a background queue is the right call — the user no longer blocks on memory processing after each turn.

The queue implementation is the idiomatic and correct shape:

• Strict FIFO via the backgroundTail = backgroundTail.then(...) promise chain — tasks run one after another, never concurrently, so ordering and the per-turn session/model context are preserved.
• Failures don't stall the chain — each task's error is caught and pushed to backgroundFailures so one bad turn doesn't block the rest, which is what you want for fire-and-forget background work.
• No lost work on exit — session_shutdown drains the tail before completing, and a drain error is recorded and rethrown rather than swallowed, so a failed final sweep still surfaces.

CI is fully green — lint/typecheck, all four test matrices, and e2e-k8s. Nice, tightly-scoped change with test coverage in pi-port.test.ts.