fix(mcp): preserve agent attribution in mempalace_checkpoint (#2023)

[mvalentsev]

What does this PR do?

Fixes #2023. mempalace_checkpoint hard-coded added_by="checkpoint" for every drawer it filed, so the originating agent's identity was dropped from drawer metadata (it surfaces in exports and provenance). The agent is already handed to the same call through diary.agent_name, so the information was available but discarded.

This adds an optional top-level added_by parameter to mempalace_checkpoint and resolves attribution in this order:

1. an explicit added_by argument, otherwise
2. the diary's agent_name (the agent filing this session), otherwise
3. the legacy "checkpoint" label.

A blank, whitespace-only, or non-string value counts as unspecified at each step, so a malformed argument defers to the next source instead of reaching sanitization. tool_add_drawer already accepts and stores added_by, so there is no storage change; the parameter is declared in the tool's input_schema, so the tools/call dispatch admits it over both the stdio and HTTP transports.

Backward compatible: calls that supply neither an explicit added_by nor a diary agent_name keep the checkpoint label. The stored value is verbatim (matching every other added_by writer); the diary index continues to lowercase agent_name separately for its case-insensitive reads.

How to test

uv run pytest tests/test_mcp_server.py -k checkpoint -v

Manual check over the MCP path:

• Call mempalace_checkpoint with one item and diary.agent_name="DeepSeek", then inspect the filed drawer: added_by is DeepSeek (was checkpoint).
• Call with an explicit added_by="alice": the drawer records alice regardless of the diary.
• Call with neither: the drawer records checkpoint.

Checklist

• Tests pass (python -m pytest tests/ -v)
• No hardcoded paths
• Linter passes (ruff check .)