fix(runtime): preserve custom model identity in OpenCode config

[Yevanchen]

Summary

• Preserve custom endpoint model IDs, including slashes, and use the same resolved model ID for the LLM proxy grant and OpenCode configuration.
• Render OpenCode's provider prefix only at the adapter boundary; remove the duplicate model parsing helpers.

Why

Selecting deepseek/deepseek-v4-flash through an openai-compatible credential previously produced an OpenCode selection of deepseek/deepseek-v4-flash, although only openai-compatible was enabled. OpenCode can then fall back to another declared model and receive a 403 because its request differs from the model authorized by the proxy.

The selection is now openai-compatible/deepseek/deepseek-v4-flash, while both the upstream model ID and grant retain deepseek/deepseek-v4-flash. Existing preset vendor-prefix normalization remains supported.

Verification

• just test-file apps/api/tests/runtime-vendor-env-vars.test.ts: 33 passed; slash-containing regression cases failed before the fix and passed afterward.
• just test-file apps/api/tests/driver-llm-proxy-route.test.ts: 34 passed, including admission of the provisioned model and rejection of altered models/protocols.
• just tc-package @mosoo/api, just lint, changed-file formatting checks, and git diff --check: passed.
• Manual: ran the pinned OpenCode 1.18.4 ACP process against a local controlled Chat Completions endpoint, with both gpt-5 and deepseek/deepseek-v4-flash declared. The old selection requested GPT and received a model-capability 403. The fixed selection requested DeepSeek throughout, executed a real bash tool writing pi-byok-ok to a temporary file, and completed the turn.
• Full gate: just check was attempted on macOS; existing driver process-supervision tests require Linux /usr/bin/setsid and fail locally. See Linux CI for the complete gate.
• Not run: production Sandbox / live upstream acceptance. The local reproduction proves this failure mechanism, but the original production request body was unavailable.

Impact

• User/API/contract changes: custom model IDs remain intact, avoiding unintended model fallback and proxy denial. No public schema or UI changes.
• Generated files / GraphQL / DB / lockfile: none.
• Env or config changes: corrected generated OpenCode model selection; no new environment variables or secrets. No Pi dependency or runtime migration.
• Risk and rollback: custom model IDs beginning with openai-compatible/ are now treated literally. Revert this commit to restore prior behavior. Deployment verification should use a new session/driver so it receives the new configuration and grant; existing generated session configuration is not rewritten.

Review

• Closest review areas: OpenCode provider selection, opaque custom model IDs, and exact model/protocol proxy authorization.
• Known trade-offs: retains the existing protocol selection and preset compatibility rules; this is a focused model-identity fix.