Route OpenCode explicit model overrides through the AI Gateway

[molebox]

Problem

An explicit model override like anthropic/claude-sonnet-5 was passed to the OpenCode CLI verbatim (--model anthropic/claude-sonnet-5). OpenCode reads the first segment as its provider id, and the generated opencode.json only configures the vercel (AI Gateway) provider — so every explicit-model run using a canonical gateway id died at session start:

{"type":"error","error":{"name":"UnknownError","data":{"message":"Unexpected server error. Check server logs for details."}}}

with observedModel: null on every task. The form OpenCode actually routes through the gateway is vercel/anthropic/claude-sonnet-5 — consistent with the package's own defaultModel shape and the README's OpenCode section. claude-code and codex were unaffected because their adapters route via gateway env/config (ANTHROPIC_BASE_URL, model_provider = \"vercel\").

Fix

Mirrors codex's host-computed model handling (openai/ prefixing):

• resolveOpenCodeModel (host-side, opencode/agent.ts) — prefixes vercel/ unless the caller already targets vercel/... or a key configured via agentOptions.extraProviders (the unreleased-models escape hatch). Host-side because only the host knows extraProviders.
• runnerExtra() — threads the resolved id to the sandbox runner as input.extra.cliModel. The judge path (eval-helper.mjs) ships extra but not agentOptions, so the resolution must happen host-side. The runner deliberately has no fallback to input.model — verbatim pass-through is the exact mis-route this fixes, and both runner entry points always ship extra.
• normalizeObservedModel (runner-side, opencode/run.mjs) — un-applies exactly the prefix the host added, so observedModel === requestedModel holds for canonical gateway ids, and a gateway substitution still surfaces as a clean gateway id (anthropic/claude-haiku-4, not vercel/anthropic/...).

Compatibility

• The README-documented vercel/{provider}/{model} form passes through verbatim — zero change for existing consumers following the docs.
• The auto-prefix path only activates on inputs that previously failed with provider-not-found, so no working consumer can regress.
• Native-default runs are untouched: CLI args are byte-identical and observations (e.g. vercel/google/gemini-3-pro-preview) pass through unchanged — no published-data drift.
• extraProviders requests are passed verbatim, preserving that workflow.
• No public API surface change (resolveOpenCodeModel has the same internal/test-only visibility as generateOpenCodeConfig) → patch changeset.
• README's OpenCode model format section updated — it previously stated the unprefixed form fails.

Testing

• 8 new unit tests covering: prefixing, idempotence on vercel/... ids, extraProviders verbatim + gateway ids still prefixed alongside them, bare ids (no /), no-verbatim-fallback in arg building, native-default arg parity, observed-model round-trip, gateway-substitution reporting, and native-default observation pass-through.
• Full suite: 259 passed, tsc + eslint clean, copy-runners build verified.