fix(daemon): let opencode choose its own port when spawning runtimes

[Aleexc12]

Summary

The daemon could register N instances but only ever run one at a time. Starting a second instance failed with:

Server exited with code 1
Warning: OPENCODE_SERVER_PASSWORD is not set; server is unsecured.
Failed to start server on port 4096

The TUI hit this when switching models, because that path triggers a fresh opencode spawn while the existing one is still holding port 4096.

Root cause

OpencodeRegistry.ensureStarted in packages/daemon/src/opencode.ts called createOpencode() with no options. @opencode-ai/sdk@1.3.13's createOpencodeServer then filled in its hardcoded default port:

// @opencode-ai/sdk v2 — createOpencodeServer
options = Object.assign(
  { hostname: "127.0.0.1", port: 4096, timeout: 5000 },
  options ?? {},
)
const args = [`serve`, `--hostname=${options.hostname}`, `--port=${options.port}`]

So every spawn asked the opencode binary to bind exactly 127.0.0.1:4096. The second concurrent spawn lost the race and exited with EADDRINUSE.

The fix

One change: createOpencode() → createOpencode({ port: 0 }).

• --port=<nonzero> — binds exactly that port; exits 1 on EADDRINUSE (no fallback).
• --port=0 — tries 4096 first, then falls back to an OS ephemeral port if 4096 is busy. Supports arbitrarily many concurrent instances.

Verification

Reproduced against @opencode-ai/sdk@1.3.13 and the bundled opencode binary:

1. With the old code (createOpencode()): Started two registered instances via ralph daemon instance start. First bound to 4096; second failed with the exact error above.
2. With the fix (createOpencode({ port: 0 })): Same two-instance drive. First bound to 4096; second bound to an ephemeral port (61264 in one run, 53157 in another). Both instances ran concurrently with no collision.
3. Sanity check on the opencode binary directly: opencode serve --port=0 alone binds 4096; running a second one concurrently lands on an ephemeral port. Confirms the fallback is inside the opencode binary, not the SDK.
4. Manual TUI validation: opening the model picker, switching models mid-conversation, and driving two real-workspace instances in parallel all succeed end-to-end.

Test plan

• Repro original failure with createOpencode() → "Failed to start server on port 4096" on second instance
• Confirm fix with createOpencode({ port: 0 }) → two instances run concurrently on distinct ports
• Confirm via direct opencode serve --port=0 invocation that fallback lives in the opencode binary
• Exercise the TUI model-switching path end-to-end (the originally reported trigger)

Files touched

• packages/daemon/src/opencode.ts — OpencodeRegistry.ensureStarted now passes { port: 0 } to createOpencode (+ a biome-driven reformat of the surrounding .then callback).