fix(agent): route memory provider tools in sequential execution path

[lance0]

Summary

The sequential tool execution path (_execute_tool_calls_sequential) was missing two dispatches that exist in the concurrent path (_invoke_tool):

1. Memory provider tool routing — tools like fact_store and fact_feedback (from holographic, hindsight, etc.) were registered in valid_tool_names and passed validation, but fell through to handle_function_call() which doesn't know about memory provider tools, causing "Unknown tool" errors.

2. Memory write bridge — when the built-in memory tool writes to MEMORY.md/USER.md, the concurrent path notifies external memory providers via on_memory_write() so they can mirror the data. The sequential path was missing this bridge, so holographic/hindsight/etc. never received built-in memory writes.

Root Cause

_invoke_tool (used by concurrent execution) has both dispatches. _execute_tool_calls_sequential (used for single tool calls and interactive tools) had its own inline dispatch that was missing them. Single tool calls are the common case for gateway/Telegram, so this affected most real-world usage.

Fix

15-line addition to _execute_tool_calls_sequential:

• elif branch routing memory provider tools through _memory_manager.handle_tool_call()
• on_memory_write() bridge after built-in memory tool execution

Both match the existing behavior in _invoke_tool.

Testing

• Confirmed memory provider registers and activates correctly (tools in valid_tool_names)
• Confirmed model generates proper fact_store tool calls (raw API test to llama-server)
• Confirmed tool validation passes (fact_store in valid_tool_names = True)
• Confirmed on_memory_write bridge works: built-in memory writes now mirror to holographic SQLite store with HRR vectors
• Tested with both Qwen3.5-27B and Gemma 4 31B via local llama-server

Fixes #4781

[fire]

I think the pull request works for me.

[lance0]

Fix merged in #4803 (v0.7.0). Closing this one — glad the diagnosis helped. 🎉