Fix proposal: retry transient Codex WebSocket failures

[valkia]

I prepared a tested fix for transient Codex WebSocket failures being surfaced as terminal PI_AI_ERROR turns instead of entering the default retry policy.

The repository currently exposes no Pull Requests endpoint for external contributors (the REST endpoint returns 404 and the GraphQL repository reports zero PRs), so I am sharing the ready-to-cherry-pick branch here:

• Branch: https://github.com/valkia/deepseek-harness/tree/fix/pi-ai-websocket-transport
• Commit: valkia@dd33e64

Problem

pi-ai flattens WebSocket failures to message text before dsh-llm-pi-ai sees them. The current classifier's standalone socket pattern does not match the compound word WebSocket, so messages such as WebSocket error and WebSocket closed 1012 fall through to PI_AI_ERROR.

The default retry policy includes TRANSPORT, but not the catch-all PI_AI_ERROR. A transient provider/network interruption therefore ends the whole turn instead of being retried.

Fix

• map generic and transient WebSocket failures to TRANSPORT
• map deterministic WebSocket close codes 1002/1003/1007/1008/1009/1010 to INVALID_REQUEST, avoiding retries of protocol, payload, policy, size, and extension failures
• leave WebSocket transport is not available in this runtime unclassified and non-retryable
• add regression cases for the observed WebSocket error and WebSocket closed 1012 messages
• update the English/Chinese package documentation and existing Agent Note

Validation

• focused convert.spec.ts: 81/81 tests, 100% coverage for src/stream.ts
• pnpm run doc-sync
• pnpm run lint
• pre-push incremental typecheck

The commit is based on upstream master at b150a551b8d465e31e418e1b2eaf5e79bbb7d28e.