fix(gateway): use filtered history length for transcript message extraction

[PercyDikec]

The transcript extraction used len(history) to find new messages from each agent turn, but history includes session_meta entries that are stripped before passing to the agent. This mismatch caused 1 message to be lost from the transcript on every turn after the first.

Changes:

• _run_agent now returns history_offset (the filtered history length actually passed to the agent) in its result dict
• The caller uses history_offset instead of len(history) for the slice
• Changed the else branch from returning all agent_messages to returning [], preventing full-history duplication when session_meta count equals agent message count

Tests

Added tests/gateway/test_transcript_offset.py with 6 tests:

• session_meta causes offset mismatch - core bug: 1 session_meta entry causes 1 message loss
• no session_meta gives same result - first turn (no session_meta) works correctly with both approaches
• multiple session_meta entries - 2 session_meta entries cause the else branch to duplicate the whole history
• system messages also filtered - system messages in history also contribute to the offset drift
• else branch returns empty list - verifies the else branch fix returns [] instead of all messages
• tool call messages preserved in filter - tool_calls and tool results pass through the filter correctly, keeping the offset accurate

Closes #394