Windows: TLS ClientHello never sent (0 bytes even to localhost), all threads deadlock — opencode hangs forever

Bug: On Windows, TLS ClientHello is never sent — process deadlocks (0 bytes even to localhost)

Environment

• OS: Windows (Git Bash shell)
• opencode version: 1.18.10 (also tested: 1.18.9, opencode-windows-x64-baseline@1.18.9 — all fail identically)
• Installed via npm i -g opencode-ai
• Provider: opencode-go (API key valid — direct curl works in 0.2s)

Symptom

Both TUI and CLI hang forever with no response and no error:

$ opencode run --model opencode-go/deepseek-v4-flash "hi"
> build · deepseek-v4-flash
   (hangs forever)

Log (~/.local/share/opencode/log/opencode.log) stops after llm runtime selected with no error:

message=loop ... step=0
message=stream providerID=opencode-go modelID=deepseek-v4-flash ... agent=title mode=primary
message="llm runtime selected" llm.runtime=ai-sdk llm.provider=opencode-go llm.model=deepseek-v4-flash
(no further log lines)

Evidence

1. TCP connects, but TLS never starts (0 bytes). While hung, netstat shows 4 ESTABLISHED connections from opencode to Cloudflare IPv6 2606:4700:78::90:0:140-143:443. A CONNECT proxy in between logs the tunnel established, then 0 bytes in both directions — not even a TLS ClientHello.
2. Same with a local HTTPS server. Pointing the provider baseURL to https://127.0.0.1:8443 (a Python HTTPS server on the same machine, self-signed cert): curl to it works (HTTP 200, request logged), but opencode connects and the server receives zero bytes — ClientHello never sent even to localhost.
3. Threads deadlock. While hung, all 30 threads are Waiting and CPU increases only ~0.08s over 3s — waiting on an event that never fires, not busy-looping.
4. Same bun version works standalone. opencode.exe embeds Bun 1.3.14 (extracted from the binary) — same as the system bun 1.3.14. With that bun, plain fetch AND import('undici').fetch to https://opencode.ai/zen/v1/models both return HTTP 200 instantly.
5. curl / node work. curl (IPv4, IPv6, direct, via proxy) and node fetch to the same endpoints all return HTTP 200 in <0.3s.

Already ruled out

• No proxy env vars; Windows system proxy disabled (ProxyEnable=0); WinHTTP direct.
• No antivirus process found; loaded DLLs are all standard system DLLs.
• API key valid (direct API call works). Config valid (opencode-go/deepseek-v4-flash).
• Not version-specific: 1.18.10, 1.18.9, and baseline build all behave identically.

Hypothesis

The AI SDK's async TLS handshake never gets scheduled — opencode's event loop is waiting on something internally (deadlock), so the TCP connect completes (synchronous) but the TLS handshake (async) never runs. This looks like a Bun/undici/ai-sdk interaction issue specific to some Windows environments.

Steps to reproduce

npm i -g opencode-ai
opencode auth login   # any provider
opencode run --model opencode-go/deepseek-v4-flash "hi"

Hangs forever, no output, no error.