[Bug] KNOWN_FAILURE_MODES §9 recommends a 5s chunkTimeout that causes "SSE read timed out" #521

What happened?

docs/KNOWN_FAILURE_MODES.md §9 ("Long Timeouts / Hanging Requests") recommends:

{ "requestTimeout": 30000, "chunkTimeout": 5000 }

chunkTimeout is the maximum gap between streamed chunks, so 5000 ms aborts any response
where the model pauses for more than 5 seconds (long reasoning, generating a large tool
call). OpenCode then shows Provider retry (attempt N): SSE read timed out.
requestTimeout is not an OpenCode provider option; the actual key is timeout.

The section also doesn't mention the SSE read timed out message, which users do hit in the
sandbox even with the kit defaults (for example on long tool calls), so there is no
documented explanation or remedy for it.

What did you expect?

The section should reflect what the usai-provider kit already configures (timeout
600000 and chunkTimeout 120000 under provider.usai.options), explain the SSE read
timeout and its automatic retry, and show how to raise chunkTimeout per project.

Steps to reproduce

1. Copy the §9 example into a project's OpenCode provider options.
2. Ask the agent for a task with a long reasoning step or a large file write.
3. Observe repeated SSE read timed out retries.

Docker version

N/A (documentation issue)

Operating System

N/A

SBX method

N/A (applies to both SBX and MSB backends)

Additional context

Upstream: anomalyco/opencode#17578 (large tool calls stalling the stream) and
anomalyco/opencode#51879 (open; changes how keep-alive comments affect the timer).