fix(local-runtime): recover stopped plugin instances - #766

Description

Fixes #765

This pull request fixes a local runtime recovery issue where plugin-daemon can continue routing requests to a local plugin instance after that instance's stdio pipes have already been closed.

When this happens, dispatch fails before the request reaches the plugin process, for example:

PluginDaemonInternalServerError: write |1: file already closed
failed to write request

The root cause is that stopped local plugin instances are removed from the runtime instance pool asynchronously. During that gap, the load balancer can still select the stale instance, and the write path can still attempt to write to its closed stdin pipe.

This PR makes the local runtime recover from that state by:

• marking PluginInstance as stopped synchronously when stdio is closed or stdout exits
• skipping stopped instances during local runtime load balancing
• evicting dead instances when stdin writes fail with closed-pipe style errors
• waking the scheduler after eviction so replacement instances can be started promptly
• retrying local write-before-dispatch failures within a small bounded window
• serializing writes to a plugin instance's stdin to avoid interleaved payloads from concurrent sessions

Type of Change

• Bug fix
• New feature
• Refactor
• Performance improvement
• Other

Essential Checklist

Testing

• I have tested the changes locally and confirmed they work as expected
• I have added unit tests where necessary and they pass successfully

Bug Fix (if applicable)

• I have used GitHub syntax to close the related issue (e.g., Fixes #123 or Closes #123)

Additional Information

Local validation:

• go test ./internal/core/local_runtime/... ./internal/core/io_tunnel/...
• go build ./cmd/server

The change is limited to local plugin runtime instance lifecycle, load balancing, stdin write handling, and write-before-dispatch retry behavior. It does not change serverless runtime behavior.