fix(lsp): prevent socket deadlock on large files with incremental sync

[JayZenith]

Issue for this PR

Closes #16115

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Patching large Python files (e.g. server_args.py in sglang, ~255KB) causes opencode to hang indefinitely on "Preparing patch...".

textDocument/didOpen and textDocument/didChange both sent the full file content over a Unix domain socket. At 255KB this overflows the ~213KB kernel socket buffer. When pyright is CPU-saturated on its initial project scan it can't drain the socket fast enough, both sides stall waiting on each other. Deadlock.

Two changes:

1. All LSP notifications are now fire-and-forget. Node.js buffers the write internally and streams to pyright as the socket drains, without blocking opencode.
2. didChange sends a line-level incremental diff instead of the full document, so subsequent patches never re-saturate the buffer.

How did you verify your code works?

Profiled both the official nightlyand this fix simultaneously using ss -xp to observe Unix socket buffer states.

Official nightly on server_args.py inside sglang:
Recv-Q: 219,461 -> pyright's buffer, not draining
Send-Q: 224,000 -> opencode blocked on write, pyright at 108% CPU.

This fix under the same conditions:
Recv-Q: 561 -> draining normally
Send-Q: 0 ->opencode not blocked
Two back-to-back patches on server_args.py completed without hanging.

Screenshots / recordings

N/A

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Found 1 potentially related PR:

#16114: fix(lsp): timeout write to prevent hang when LSP server stdin is full

• fix(lsp): timeout write to prevent hang when LSP server stdin is full #16114

This PR appears to be addressing a related issue (socket buffer saturation causing hangs) but uses a different approach (write timeouts). Since PR #16538 (the current PR) closes issue #16115 and provides a more comprehensive fix with incremental diffs and fire-and-forget notifications, #16114 may be an earlier attempt or alternative solution to the same underlying problem.

[github-actions]

Thanks for updating your PR! It now meets our contributing guidelines. 👍

[JayZenith]

@thdxr found this while working on sglang, took a while to track down but the socket profiling made it pretty clear what was happening. lmk if you have any questions.