feat(mcp): add mempalace_delete_by_source bulk-cleanup tool (#1722)

[arnoldwender]

What and Why

#1722 reports benchmark/eval data (ShareGPT dumps, results_mempal_*.jsonl, language config JSON) getting mined into the same wing as real user data — in the reporter's palace, 55k of 70k embeddings (78%) were noise, making semantic search return yoga-retreat conversations for queries about real clients. The only recourse today is hand-rolling DELETE against chroma.sqlite3.

This adds the immediate ask from that issue: an MCP tool to bulk-remove drawers by their source_file. (Requests #2 ingestion-time validation and #3 namespace isolation are deliberately left out — separate design calls.)

Change Summary

• mcp_server.py — new tool_delete_by_source(source_file, dry_run=True) + mempalace_delete_by_source TOOLS entry + header doc line.
  • Matching is exact on source_file and pushed to the backend via delete(where={"source_file": ...}) — the same idiom miner.py, diary_ingest.py, format_miner.py and palace.py already use. No client-side id list, so the SQLite "too many variables" limit cannot be hit regardless of how many drawers share the source.
  • Dry-run counting uses the existing _fetch_all_metadata paginator, so it stays accurate past the 10k get() truncation (the reporter had 55k matches).
  • Defaults to a dry run: returns match_count + a distinct (wing, room) sample so the caller can confirm blast radius before committing. dry_run=false commits (irreversible). Absent source is an idempotent no-op, not an error. Commit path writes a WAL audit entry and invalidates the metadata cache.
• tests/test_mcp_server.py — TestDeleteBySource: dry-run-doesn't-delete, commit-deletes-only-matching-source, idempotent-no-match, empty-source-rejected, registered-and-dispatchable (defaults to dry-run through handle_request).
• website/reference/mcp-tools.md — reference section (keeps test_readme_claims doc-sync green).

Test Plan

• ruff check . + ruff format --check clean (ruff 0.15.14, matching CI)
• pytest tests/ --ignore=tests/benchmarks — 2435 passed, 5 skipped
• New TestDeleteBySource — 5 passed
• Manually traced: dry-run on a seeded palace reports the right count and leaves all drawers; commit removes only the matching source

Related to #1722

[gemini-code-assist]

Code Review

This pull request introduces a new tool, mempalace_delete_by_source, which allows bulk deletion of drawers mined from a specific source file, helping clean up accidental test or benchmark data. It includes a dry-run mode by default, comprehensive unit tests, and updated documentation. The review feedback suggests two improvements: defensively validating that source_file is a string and stripping lone surrogates for consistency, and defaulting missing wing or room metadata to empty strings in the sample preview.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

To prevent potential AttributeError if a non-string is passed to source_file, we should defensively check that it is a string. Additionally, we should strip lone surrogates from source_file using strip_lone_surrogates to ensure exact matching consistency with how source_file is sanitized and stored during ingestion (e.g., in tool_add_drawer).

Suggested change

	if not source_file or not source_file.strip():
	return {"success": False, "error": "source_file must be a non-empty string"}
	if not isinstance(source_file, str) or not source_file.strip():
	return {"success": False, "error": "source_file must be a non-empty string"}
	source_file = strip_lone_surrogates(source_file)

[gemini-code-assist]

If wing or room metadata is missing or None, meta.get() will return None. To maintain consistency with other tools in this file (which default missing wing and room to ""), we should default them to "" before adding them to the sample list.

Suggested change

	for meta in metas:
	meta = _safe_meta(meta)
	key = (meta.get("wing"), meta.get("room"))
	if key in seen:
	continue
	seen.add(key)
	sample.append({"wing": meta.get("wing"), "room": meta.get("room")})
	if len(sample) >= 5:
	break
	for meta in metas:
	meta = _safe_meta(meta)
	wing = meta.get("wing", "")
	room = meta.get("room", "")
	key = (wing, room)
	if key in seen:
	continue
	seen.add(key)
	sample.append({"wing": wing, "room": room})
	if len(sample) >= 5:
	break

[arnoldwender]

Thanks for the review @gemini-code-assist — both suggestions addressed in 402ee73:

1. source_file normalization + type guard (mcp_server.py): good catch on the consistency gap. tool_add_drawer strips lone surrogates from source_file before storing, so exact matching would silently miss rows mined from non-ASCII paths arriving via a cp1252 stdin (MCP server fails on Cyrillic/non-ASCII content on Windows due to default cp1252 stdin encoding #1488). The handler now applies strip_lone_surrogates to the input before building the where filter, and guards with isinstance(source_file, str) so a non-string returns the clean error instead of AttributeError. Added a test that adds a drawer with a lone-surrogate path and confirms deletion still matches, plus a non-string-rejection test.
2. Default wing/room to "" in the sample — applied, consistent with the rest of the file.

CI re-running on 402ee73. Ready for re-review.

[gemini-code-assist]

Code Review

This pull request introduces a new MCP tool, mempalace_delete_by_source, which enables bulk deletion of drawers based on an exact match of their source_file metadata. The tool includes a default dry-run mode to preview matches and a commit mode to execute the deletion. It also incorporates surrogate normalization to ensure consistent matching. Accompanying tests and documentation updates have been added to verify and describe the new functionality. There are no review comments, and the implementation looks solid, so I have no additional feedback to provide.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[arnoldwender]

Rebased on latest develop (post-3.4.1, +84 commits). Conflict resolution only — no logic changes to the tool:

• mcp_server.py: the new tool_mine/_capture_fd_stdout from Feature request: add mempalace_mine tool to MCP server #1662 landed in the same slot between tool_delete_drawer and tool_sync; kept both functions and both TOOLS dict entries (mempalace_mine + mempalace_delete_by_source).
• website/reference/mcp-tools.md: kept both reference sections; intro count is now 34 (33 after Feature request: add mempalace_mine tool to MCP server #1662 merged, +1 for this tool). test_readme_claims parity verified — 34 dict keys / 34 doc sections, no orphans.

Full suite green locally on .venv (2861 passed, 5 skipped); ruff check/format clean. Ready for re-review.