fix(reflect): reject serialized parameter tails in done answers

[Alan5168]

Summary

Reject a parsed done tool call when its answer contains a serialized sibling parameter tail such as </answer><parameter name="memory_ids">....

Some provider responses can return a valid tool call while blending XML tool serialization into the JSON answer value. Because a tool call exists, the no-tool-call guard added in #3013 does not run, and the corrupted tail gets persisted verbatim into mental_models.content.

This validates the exact answer/sibling-parameter boundary against the declared done() fields and raises ReflectToolCallError (matching the existing fail-loud contract). It does not truncate or silently clean the answer, and it does not touch #3013's no-tool-call guard.

Tests

• Added test_done_tool_rejects_serialized_parameter_tail, a regression test that is red on current main and green with the fix (verified locally by reverting just the production fix and re-running).
• 40 passed across the related pure unit tests in test_reflect_agent.py (excluding RealLLM/Integration/DirectiveLeakageOnEmptyBank, which need external providers/deps not available here).
• ruff check and ruff format --check pass for all changed Python files.

Closes #3048

[nicoloboschi]

seems like a model issue, let's not open the pandora's box