On a complex plan (building a large nested structure), the model can
hallucinate a tool name more than once — emitting a primitive node type
(heading/text/image) as the "tool" — before it re-reads the tool list and
picks the real one. A single retry dead-ended those turns with a cryptic
"Tool <x> is not available" surfaced straight to the user. Bump the
self-correction cap from 1 to 2 (still bounded); the feedback already lists
the valid tool names.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>