fix(agent): handle duplicate tool call IDs in OpenAI-compatible transcripts

[badgerbees]

The Issue

Several "OpenAI-compatible" APIs (including OpenRouter, vLLM, and certain local models) strictly enforce unique tool_call_ids across the entire conversation history. If a transcript contains duplicate IDs, the provider returns an HTTP 400 Bad Request, effectively breaking the session. This happens in two ways:

1. Live collisions: The model returns two tool calls with the same ID in a single response (common in DeepSeek and certain open-source models).
2. Historical reuse: The model happens to reuse an ID (e.g., call_abc) in a later turn that was already used in an earlier turn.

The Fix

This PR implements an "occurrence-aware" ID sanitization strategy to ensure transcript-level uniqueness without losing data.

1. Occurrence-Aware Sanitization: Updated sanitizeHistory in internal/agent/loop_history.go to track how many times a raw ID has been seen in the transcript. Duplicate IDs are elegantly rewritten (e.g., id , id_1, id_2) while maintaining the correct pairing between assistant calls and tool results via encounter-order queues.
2. Live ID Normalization: Added normalizeToolCallIDs in internal/agent/loop.go. This catches and fixes collisions immediately as they are returned by the LLM, ensuring that new messages join the history with unique IDs from the start.
3. Deep-Copy Safety: Added deep-copy logic for ToolCalls slices during sanitization to avoid shared-memory side effects on the original session history object.
4. Anthropic/Thinking Safe: Added a critical guard that skips ID rewriting if RawAssistantContent is present. This is essential for Anthropic Claude models (or any model using raw content passback) to avoid breaking the link between the provider's internal thinking blocks and tool results.