fix(schema): publish config schema that accepts V1 and V2 shapes - #55524

Summary

https://opencode.ai/config.json is the $schema used by every opencode.json(c), and both V1 and V2 read that file. Today it's a static V1 file that was frozen into packages/console/app/public when packages/opencode was removed from dev. As a result, editors flag valid V2 configs as errors: permissions, agents, providers, mcp.servers, skills: [], update, warming, and others.

This PR builds the file during the deploy, so it now matches how V2 parses config:

• V2 shape: generated from Config.Info in @opencode/schema/config.
• V1 shape: taken from a frozen snapshot of the last V1 schema (packages/schema/script/config-v1.json, moved here and refreshed from today's v1 branch; only the autoupdate description changed). Fields only V1 understands are marked deprecated. This matches ConfigNormalize, which accepts V1 and V2 fields mixed in one file.
• Keys both versions share: they keep the V2 shape. Shared objects also accept their V1 properties (mcp legacy server entries, compaction.reserved, experimental.mcp_timeout, ...). skills changed type, so it accepts either shape.
• Schema structure: the root is a flat object schema rather than a root $ref. allowComments and allowTrailingCommas stay.

Deployment

• Console app: its build runs the generator after vite build and writes .output/public/config.json. This is the console's build script before 5caf1f9f6f, now pointed at the new generator. The dev and production deploys serve it at opencode.ai/config.json. The v1 branch only deploys app.opencode.ai, so it never touched this file.
• Docs site: the packages/web build hook (docs.opencode.ai/config.json) now runs the generator instead of copying the static file.
• Staying current: every deploy regenerates the file from Config.Info, so future V2 config changes need no manual step.

Other changes

• V2 descriptions: top-level V2 descriptions are copied from the type-side document, because fields declared as optional lose their annotations on the encoded side.
• Skill doc: the OpenCode skill no longer says the schema "may describe V1".

Known gaps (follow-ups)

• Nested descriptions: nested V2 fields still have no descriptions; most of these schemas are annotated after optional. Where V1 had a matching field, its description is used instead.
• Legacy MCP names: a legacy MCP server literally named servers or timeout validates against the V2 shape.

Validation

Done locally; no test is added to the PR. The deploy builds fail loudly if the generator throws.

• Real builds: bun run build in packages/console/app and packages/web both emit config.json, byte-identical to the generator output.
• No regressions against the live schema: all 2,143 opencode.json examples in the V1 (all locales) and V2 docs, plus 680 random V1-valid configs from a fuzzer. Nothing the live schema accepts is rejected.
• V2 coverage: 400 random values generated from Config.Info, encoded, all accepted. The samples cover every top-level key.
• Agreement with the V2 runtime: about 50 hand-written V1/V2/mixed edge cases checked against ConfigNormalize. The only disagreements are the known gaps above, plus unknown keys, which the schema rejects and the runtime ignores.
• Future changes:
  • a new V2 field (top-level or nested) appears automatically;
  • a V2 key that reuses a V1 name with a new type becomes anyOf;
  • a V1/V2 definition-name clash fails the build;
  • converting Config.Info from a Class to a Struct still generates identical output.
• bun run check passes.

Issues

Fixes #53398
Fixes #50724
Fixes #48812
Fixes #43748
Fixes #41014
Fixes #46248

Related: #53744, #52177