fix(opencode): omit tool_choice for zai providers when tools are empty

[AceRothstein71]

What

Injects a sanitizing fetch wrapper into the options.fetch of the four Z.AI/Zhipu provider entries (zai, zai-coding-plan, zhipuai, zhipuai-coding-plan) during the plugin config hook, so that tool_choice is dropped from chat-completion request bodies when the effective tools list is empty/absent or when tool_choice names a function not present in tools. Valid requests and SSE responses pass through byte-identical.

New module: packages/omo-opencode/src/features/zai-tool-choice-guard/ wired into config-handler.ts next to applyOpenGatewayProviderConfig.

Why

The bundled @ai-sdk/openai-compatible provider serializes tool_choice unconditionally. Z.AI/GLM endpoints reject such bodies with HTTP 400 code 1210 (AI_APICallError: API 调用参数有误), so Atlas/Sisyphus turns that force tool_choice:"required" fail intermittently on zai-coding-plan/glm-5.2 whenever MCP/provider-defined tools leave the function-tool list empty (#6753).

OMO cannot patch the vendored SDK, but opencode's provider pipeline makes an in-process fix possible (verified against opencode dev source and the installed 1.18.16 binary):

• plugin config() hooks run before cfg.provider is read (packages/opencode/src/provider/provider.ts ~1420-1424)
• config provider entries are deep-merged over the models.dev catalog entry, preserving its models (~1462-1471, re-applied ~1628-1634)
• the openai-compatible loader passes options.fetch through to createOpenAICompatible

So the guard covers every request path for those providers without a local proxy, and stays inert for all other providers.

Verified

• Failing-first: regression tests written before the implementation; red confirmed (Cannot find module './index'), then green after implementing.
• New co-located suite (19 given/when/then tests): tool_choice stripped for absent/empty tools and missing named function; byte-identical passthrough for valid choices and choice-less bodies; non-JSON body verbatim; SSE Response passthrough; user fetch composed not replaced; idempotent brand prevents double-wrap; Request-object inputs handled.
• Wiring tests: creates missing zai entries with options.fetch, preserves existing models/options on them, leaves anthropic/github-copilot untouched, second pass is a no-op.
• Scoped suites: guard + config-handler.test.ts = 71 pass / 0 fail; wider net (plugin-handlers + opengateway-provider) = 266 pass / 0 fail.
• bun run typecheck exit 0 (tsgo root + script + packages).
• Evidence recorded under .omo/evidence/20260824-6753-glm52-tool-choice/ (WHAT TESTED / OBSERVED / WHY ENOUGH / OMITTED), committed with git add -f.

Risk

Low and contained: only zai-family provider entries gain an options.fetch wrapper; every other provider is untouched. If a future opencode stops honoring options.fetch, the wrapper becomes inert rather than breaking requests. The upstream SDK-level fix remains the complete remedy for non-zai openai-compatible providers with the same validation behavior.

Fixes #6753

---

Summary by cubic

Drops invalid tool_choice from Z.AI/Zhipu chat-completion requests to prevent GLM 400 (code 1210) when tools are empty or missing. Previously we always sent tool_choice from @ai-sdk/openai-compatible; now a sanitizing fetch runs only for zai-family providers, leaving valid requests and streaming unchanged.

• Injects applyZaiToolChoiceGuard in config-handler.ts to wrap options.fetch for zai, zai-coding-plan, zhipuai, and zhipuai-coding-plan. User-defined fetch is wrapped, not replaced; wrapper is idempotent and passes non-JSON bodies and SSE through unchanged.
• Adds packages/omo-opencode/src/features/zai-tool-choice-guard/ (sanitizer + tests). Only these providers are mutated; existing models/options are preserved; other providers are untouched. Verified by scoped tests (71 pass), broader suites (266 pass), and repo typecheck.

^{Written for commit fc943f2. Summary will update on new commits.}