fix: merge split assistant messages to prevent DeepSeek 400 on role alternation

[woaini30050]

Description

When Hermes builds the API request payload for the next turn, some code paths (streaming accumulation, provider adapters, session replay) can split a single assistant response that contains both content (text reasoning) and tool_calls into two consecutive assistant messages:

msg[N]:   {"role": "assistant", "content": "Let me search..."}
msg[N+1]: {"role": "assistant", "tool_calls": [...]}

Strict providers like DeepSeek v4 Flash enforce strict assistant ↔ tool alternation and reject two consecutive assistant messages with HTTP 400:

Error code: 400 - Messages with role 'tool' must be a response to 
a preceding message with 'tool_calls'

The DB storage is correct (1 row with both content + tool_calls), but the API payload incorrectly has 2 messages.

Fix

Adds merge_split_assistant_messages() in agent.agent_runtime_helpers.py that merges adjacent assistant messages where:

• msg[N]: has content but NO tool_calls
• msg[N+1]: has tool_calls but NO content

→ Result: a single {"role": "assistant", "content": "...", "tool_calls": [...]}

The merge runs inside sanitize_api_messages(), which executes before every API call — covering both the main conversation loop and the summary/max-iterations path. This is a defense-in-depth approach that catches the split regardless of where it originates.

Safety (what is NOT merged)

• Two content-only assistant messages → left alone (legitimate flow)
• Two tool_calls-only assistant messages → left alone (parallel calls)
• Already-correct messages with both content + tool_calls → left alone

Testing

• Unit tests pass for all 7 edge cases

Related Issues

Closes #49147

[teknium1]

Merged via PR #55603 (commit cbe397ef458bc715ce82451e722ee0e45a67687e) — your fix is preserved via a Co-authored-by trailer.

We consolidated this with the sibling reports (#29148, #49147) and the other open PRs (#49162, #34510) into one pass in repair_message_sequence() that covers both adjacency shapes: tool_calls-split-across-two-assistants (your case) and content-only → tool_calls-only. Added a guard so a tool result between two assistants doesn't collapse distinct rounds, and an exemption for Codex Responses interim turns. Thanks for the fix — closing as merged.