fix(pi-subagent): remove top-level allOf from tool input_schema

Anthropic rejects oneOf/allOf/anyOf at the top level of a tool's
input_schema, and every Anthropic-family frontend on OpenRouter (Azure,
Bedrock, Google) surfaces the same 400, making the subagent tool unusable
with Claude models (e.g. Fable).

The allOf only enforced profile-vs-agent on single tasks, which is already
enforced at runtime by profileSelectionError; that check now runs
unconditionally instead of piggybacking on schema rejection.
assertObjectToolSchema now also rejects top-level combinators so this
cannot regress.