[darinkishore]

What

Stacked on top of PR #81. Fixes and improvements found during adversarial review.

LM layer hardening

• Tool-loop reasoning preservation: execute_tool_loop final text turn now preserves full_content (reasoning blocks) instead of dropping them via Message::assistant(&text)
• Auto+no-tools content preservation: when tools are requested but unavailable, preserves actual provider content instead of injecting synthetic warning text
• Call API collapse: call_with_tool_loop_mode collapsed to call(messages, tools, mode)
• Panic → Result: request history construction now returns errors instead of panicking
• pub(crate) boundaries: internal rig conversion methods (to_rig_message, system_prompt, to_rig_chat_history) narrowed to crate-internal

Predicted return shape unification

• Chat added as peer field on Predicted alongside output and metadata
• Predicted::chat() accessor, into_parts() returns (O, CallMetadata, Chat)
• Predict::forward returns Predicted (not tuple) — single return shape
• All call sites updated: optimizers, ReAct, module_ext, examples, tests

Doc updates

• core/mod.rs, core/module.rs, examples/01-simple.rs updated to document three-channel Predicted (output + metadata + chat)

Verification

• cargo fmt --check ✅
• cargo clippy --all-targets ✅
• cargo test --all ✅