refactor(llm): colocate per-type factories on their namespaces

[kitlangton]

Why

packages/llm already follows a clean pattern everywhere else: per-type factories live on the type's own namespace (Message.user, Message.assistant, Message.tool, Message.make, Message.text, SystemPart.make, ToolDefinition.make, ToolCallPart.make, ToolResultPart.make, ToolChoice.make, ToolChoice.named, GenerationOptions.make, HttpOptions.make, ModelRef.update, LLMRequest.update).

src/llm.ts was also re-exporting those same factories at the top level as LLM.user, LLM.assistant, LLM.toolMessage, LLM.toolCall, LLM.toolResult, LLM.toolDefinition, LLM.toolChoice, LLM.toolChoiceName, LLM.generation, LLM.system, LLM.message, LLM.text. Two construction paths to the same value is one too many — this isn't matching the AI SDK (which has no factories at all), it's matching ourselves.

The top-level LLM namespace stays focused on the request-shaped call API: LLM.request, LLM.generate, LLM.stream, LLM.model, LLM.limits, LLM.stepCountIs, LLM.requestInput, LLM.updateRequest, LLM.generateObject.

What changed

• Deleted 12 alias re-exports from packages/llm/src/llm.ts: text, system, message, user, assistant, toolDefinition, toolCall, toolResult, toolMessage, toolChoiceName, toolChoice, generation.
• Rewrote LLM.request's internal use of those aliases to call the canonical forms directly (Message.make, Message.user, ToolDefinition.make, ToolChoice.make, GenerationOptions.make).
• Rewrote every call site in packages/llm/test/** to the colocated form (one provider test file plus the shared recorded-scenarios.ts and the unit tests).
• Updated the tool-loop example in packages/llm/AGENTS.md to match.
• Added a "Conventions" section to packages/llm/AGENTS.md documenting the rule.

No call sites existed outside packages/llm/.

Migration

Old	New
LLM.user(x)	Message.user(x)
LLM.assistant(x)	Message.assistant(x)
LLM.toolMessage(x)	Message.tool(x)
LLM.message(x)	Message.make(x)
LLM.text(x)	Message.text(x)
LLM.system(x)	SystemPart.make(x)
LLM.toolDefinition(x)	ToolDefinition.make(x)
LLM.toolCall(x)	ToolCallPart.make(x)
LLM.toolResult(x)	ToolResultPart.make(x)
LLM.toolChoice(x)	ToolChoice.make(x)
LLM.toolChoiceName(x)	ToolChoice.named(x)
LLM.generation(x)	GenerationOptions.make(x)

All types are re-exported from @opencode-ai/llm's barrel via export * from "./schema".

Test plan

• bun run --cwd packages/llm typecheck — clean
• bun run --cwd packages/llm test — 174 pass / 31 skip / 0 fail
• bun run --cwd packages/opencode typecheck — clean
• bun run --cwd packages/opencode test — only pre-existing unrelated failures (Worktree > list > uses parent folder name... and session.created event...); neither test references LLM.*