feat(sdk/llm): layered per-model config + caps via optional store interfaces

[lIang70]

Summary

Closes #24.

DefaultResolver previously consumed a single ProviderConfigStore and silently dropped any per-model overrides users wrote — e.g. extra.caps stored on a model:openai/gpt-4o row never reached CapsMiddleware. This PR adds two opt-in store extensions, one strong-typed override field, and rewrites createLLM so per-model config rounds-trip end-to-end.

New surface

Symbol	Purpose
ModelConfigStore.GetModelConfig(ctx, provider, model)	Optional. Per-model {Caps, Extra} overrides; errdefs.NotFound is silently ignored, any other error fails Resolve.
DefaultModelStore.GetDefaultModel(ctx)	Optional. Typed default-model pointer, replacing the magic "global_default" provider row.
ProviderConfig.Caps	Strong-typed provider-wide caps that no longer require stuffing into the untyped Config map.
ModelConfig{Provider, Model, Caps, Extra}	New override struct returned by ModelConfigStore.
DefaultModelRef{Provider, Model}	New typed return for DefaultModelStore.
WithExtraCaps(caps)	Renamed from WithModelCaps (made the "additive" semantics explicit).

Caps merge order

createLLM now OR-merges five layers — any layer disabling a capability wins:

1. ProviderRegistry.LookupModelCaps(provider, model) — static catalog
2. ProviderConfig.Caps — provider-wide user config
3. capsFromConfig(merged) — legacy Config["caps"] (deprecated, still honored)
4. ModelConfig.Caps — per-model user config
5. resolver.extraCaps — WithExtraCaps/WithModelCaps

ModelConfig.Extra is shallow-merged (per-model keys overwrite provider keys) before being handed to the provider factory.

Default-model lookup order

Resolve("") now tries:

1. DefaultModelStore.GetDefaultModel (preferred)
2. Legacy "global_default" provider row (deprecated)
3. WithFallbackModel

Backward compatibility

• Provider-only stores (e.g. animus) keep working unchanged — Resolve only type-asserts the optional interfaces.
• The legacy "global_default" lookup still runs when DefaultModelStore is absent or returns NotFound.
• capsFromConfig still parses the legacy Config["caps"] sub-object.
• WithModelCaps is preserved as a thin alias forwarding to WithExtraCaps.

Deprecations (all tagged for removal in v0.2.0)

This PR also audits the rest of sdk/ and ensures every Deprecated: tag explicitly states removal in v0.2.0:

• llm.GlobalDefaultProvider — implement DefaultModelStore instead
• llm.WithModelCaps — use WithExtraCaps
• llm.capsFromConfig / Config["caps"] — use ProviderConfig.Caps / ModelConfig.Caps
• kanban.TaskBoard, kanban.NewTaskBoard — use Board / NewBoard
• (*kanban.Scheduler).SetKanban — use WithScheduler
• compiler.ValidateGraphDef — call def.Validate() directly
• telemetry.WithLogExporter / WithLogConsole / WithLogMinSeverity (already tagged before this PR)

sdkx/ has no Deprecated: symbols and is unchanged.

Test API hygiene (commit 2)

Resolver tests no longer construct &defaultResolver{...} directly or reference deprecated symbols. They go through DefaultResolver(...) + WithFallbackModel / WithExtraCaps, plus a single test-only newResolverWithRegistry helper that swaps the registry on a real DefaultResolver — the only field tests need to override that the public API doesn't expose.

Tests covering the legacy magic key are renamed with a LegacyMagicKey_ prefix and a comment pointing at the v0.2.0 removal, so they're easy to delete together with the constant.

Test plan

• cd sdk && go test ./... — all green
• cd sdkx && go test ./... — all green (no consumer changes needed; SDKX doesn't touch llm.ProviderConfig / DefaultResolver)
• cd plugin && go test ./... — all green
• go build ./... && go vet ./... per module — clean
• New tests in sdk/llm/resolver_layered_test.go cover:
  • per-model Extra shallow-merge over provider config
  • ModelConfigStore NotFound vs other-error semantics
  • DefaultModelStore preferred / falls back to legacy magic key / falls back to WithFallbackModel
  • 4-layer caps OR merge (catalog × provider × model × resolver-extra)
  • legacy Config["caps"] still respected (deprecation regression)
  • provider-only stores keep working (animus regression)
• Auto-tag workflow will cut sdk/v0.1.11; cascade PR to bump sdkx/go.mod will follow automatically.

Notes for reviewers

• unwrapCaps exists to avoid double-wrapping CapsMiddleware: NewFromConfig already wraps the instance once with (catalog ⊕ legacy Config["caps"]), and the resolver re-wraps with the full 5-layer composition. Without unwrapCaps the caller would see two stacked wrappers — functionally correct but harder to reason about and slightly less efficient.
• mergeCaps was changed from (a, b) to variadic (...) to make the 5-layer composition in createLLM read top-to-bottom without nesting.
• migrateVarsMessages in sdk/workflow/board.go is left untouched: it's unexported (Go tooling never treats it as a deprecation) and its "v2 migration window" refers to the board-snapshot data format, not the SDK module version.

Made with Cursor