fix(opencode-plugin): decouple flush/threshold-commit from session.idle and make state save process-safe

[TE-N-ElvenWang]

Problem

The opencode plugin's message flush and threshold-commit path
(flushPendingMessages → maybeCommitByThreshold) is wired exclusively to the
session.idle event. OpenCode emits session.idle only when an assistant turn
fully completes (no queued work). During long multi-step agent turns
(back-to-back tool calls), session.idle is never emitted, so:

1. Messages accumulate locally as captured: false but are never POSTed.
2. pending_tokens on the server never grows.
3. The configured commitTokenThreshold is never reached, so threshold-based
   commits never fire.
4. Persistence only happens at plugin dispose (shutdown) via flushAll.

Additionally, saveState uses a single fixed temp file (${statePath}.tmp).
When two opencode processes run concurrently, both race on the same .tmp,
causing ENOENT rename failures and cross-process state clobbering.

Fix

Fix A: Process-safe temp file

Use a per-PID temp file: ${statePath}.${process.pid}.tmp — eliminates
cross-process ENOENT rename races.

Fix B: Periodic flush timer

Add a setInterval-based periodic flush independent of session.idle:

• Default 60s interval (configurable via periodicFlushIntervalMs, floor 10s)
• Scans sessions for uncaptured messages, calls flushSession(..., { reason: 'periodic' })
• Starts in init(), cleared in flushAll(), .unref()'d

This decouples the send + threshold-check path from session.idle, so messages
flush and threshold commits fire even during long multi-step agent turns.

Verification

Tested in production: Periodic flush timer started appears at startup;
pending_tokens accumulates during long turns; threshold commit fires at
≥20000 without dispose; zero ENOENT races under dual-process concurrency.

[TE-N-ElvenWang]

虽然是AI写的patch啊，但是问题是opencode里用openviking的话会导致上传信息不准确的问题。尤其是多个opencode进程一起开着的时候。

[TE-N-ElvenWang]

all patches are refined by open-code-review tools with claude opus 4.8.