feat(anthropic): add AnthropicToolIdSanitizationMiddleware

[Mason Daugherty (mdrxy)]

Add AnthropicToolIdSanitizationMiddleware so cross-provider conversation histories replay cleanly against Anthropic. Anthropic enforces tool_use.id matching ^[a-zA-Z0-9_-]+$, but providers like Kimi-K2 emit IDs of the form functions.<name>:<idx> — replaying such a thread raises a 400. The middleware rewrites only the outgoing ModelRequest, leaving graph state and persisted checkpoints untouched, so it composes cleanly with HITL resume.

Changes

• Build a deterministic bad_id -> safe_id map per request: substitute illegal characters with _, then fall back to a truncated sha256 suffix (with a counter for further collisions) when the base would shadow an already-valid id or another rewritten id (_make_safe_id).
• Rewrite IDs across every Anthropic-supported shape: AIMessage.tool_calls[*].id, AIMessage.content blocks (tool_use, server_tool_use, mcp_tool_use), ToolMessage.tool_call_id, and *_tool_result content blocks (tool_result, web_search_tool_result, code_execution_tool_result, bash_tool_result, mcp_tool_result). Suffix matching on *_tool_use / *_tool_result keeps future block kinds covered.
• Drift-correct position-paired tool_calls[i] and the i-th client tool_use block onto one canonical id, then propagate that as an alias to the immediately-following ToolMessage results so cross-message linkage survives even when the input list disagreed.
• Provider-gate via _should_run: non-ChatAnthropic models follow the unsupported_model_behavior knob — 'ignore' (default; logs a debug breadcrumb so the bypass is observable), 'warn', or 'raise'. The constructor runtime-validates the literal so typos raise ValueError instead of silently degrading to ignore semantics. The langchain import is guarded with a narrowed ModuleNotFoundError so genuine import-time failures inside langchain.agents.middleware aren't masked by a misleading "install langchain" message.
• Sync wrap_model_call and async awrap_model_call share the same sanitization path.

[Eugene Yurtsev (eyurtsev)]

Any chance we want a more general version of this middleware in LangChain?

Something like StandardizationMiddleware and it'll handle some logic for standardizing things across providers?

OK if we think we want to keep it specific to Anthropic