Summary

• Extended-thinking Claude models (Opus 4.8+) sign each thinking block against its original position in the response. When Hermes separated blocks into reasoning_details and tool_calls and replayed them as [all thinking][text][all tool_use], the positions no longer matched the signatures, causing a non-retryable HTTP 400 on every subsequent turn -- an unrecoverable crash-loop.
• normalize_response() in agent/transports/anthropic.py now records the original interleaved block order in provider_data['_anthropic_interleaved_order'] (only when there are multiple thinking blocks alongside tool_use blocks, i.e. when reordering would actually invalidate signatures).
• _convert_assistant_message() in agent/anthropic_adapter.py uses this order when present to reconstruct the exact original sequence (e.g. [thinking, tool_use, thinking, tool_use, ...]), with a safe fallback to the original behaviour for messages that lack the field.

Files changed

• agent/transports/anthropic.py -- track interleaved order during normalize_response()
• agent/transports/types.py -- expose anthropic_interleaved_order property on NormalizedResponse
• agent/chat_completion_helpers.py -- propagate _anthropic_interleaved_order into the stored message dict
• agent/anthropic_adapter.py -- use interleaved order in _convert_assistant_message() to reconstruct faithful block sequence

Test plan

• Multi-turn conversation with an extended-thinking Anthropic model (Opus 4.8+) that produces both thinking blocks and tool_use blocks in the same response -- subsequent turns should not receive HTTP 400
• Single thinking block with no tool calls -- falls back to original path, no regression
• Tool calls with no thinking blocks -- falls back to original path, no regression
• Kimi/DeepSeek reasoning_content injection path -- guard condition (_already_has_thinking) still prevents duplicate thinking blocks

Fixes #35975.