refactor: flatten provider settings across core and the AI package

[rekram1-node]

Problem

opencode settings are one flat bag everywhere (config, models.dev, catalog, variants). Native @opencode/ai packages read request options from a nested providerOptions object inside model(id, settings). Only the legacy aisdk: mapping path lifted flat keys into that object, with a hand-maintained whitelist per npm package.

So the same intent had multiple spellings and which one worked depended on the package prefix:

config	aisdk:@ai-sdk/openai	@opencode/ai/providers/openai
settings.reasoningEffort	works	silently dropped → route falls back to medium
settings.providerOptions.reasoningEffort	double-nested, dropped	works

The first failure also hit models.dev variants: overriding a provider's package to a native one in config made every generated reasoning variant a no-op.

Change

There is no providerOptions concept at any boundary opencode owns. Settings are flat in config, in the catalog, and at the AI package entrypoint.

@opencode/ai

model(modelID, settings) takes one flat object. Each of the 37 entrypoints destructures the connection keys it declares (apiKey, baseURL, region, resourceName, …) and passes the rest to the route as providerOptions. Settings types become ProviderPackage.Settings & <ProtocolOptionsInput> & { connection keys }, so autocomplete and type-checking are unchanged and the auth-exclusivity type tests still hold. Internal configure(...) / Config shapes are untouched; only the package entrypoint is flat. Package guide updated.

packages/core

• Provider.nativeSettings drops opencode transport keys (timeout, chunkTimeout, fetch), which only the AI SDK runtime consumes. That is all; core knows no package key names.
• AISDKNative.map is purely the legacy translator: package mapping plus the spellings that differ between AI SDK and native packages (Bedrock connection keys and Converse body fields, OpenRouter attribution headers, Azure/Mantle chat-vs-responses selection). The legacy AI SDK shape is decoded once with a lenient schema at the top of map, replacing 31 per-field isRecord/typeof checks. ~400 → ~200 lines.
• ModelResolver applies nativeSettings once on the native path and loads whatever package the catalog names.

Tests

Existing tests updated to the flat shape (packages/ai/test/provider-package.test.ts, native-providers, meta, groq; core aisdk-native, model-resolver). One small nativeSettings case in provider.test.ts. packages/ai: 1300 pass. packages/core affected suites: 385 pass.

Behavior notes

• A wrongly-typed request option (e.g. store: "yes") makes the OpenAI-family option schema decode return nothing for that request. The old aisdk: whitelists masked this; it was already native-path behavior. Making that decode lenient per field is a separate @opencode/ai change.
• OpenRouter forwards unknown request options to its request body. Connection keys are destructured out before the rest becomes providerOptions, so apiKey/baseURL/headers/body never reach it.

Follow-up (not in this PR)

Register native packages and native-vocabulary variants at catalog ingest (models.dev, config) so the catalog itself is native and the legacy translator runs only at those entry points; add best-guess default variants for config-defined models with none.