feat(mcp): add get_checksum tool for document integrity verification

[jaymeklein]

What

Adds a new MCP tool get_checksum that computes a fresh SHA-256 over a documents stored bytes on every call, returning the digest alongside the documents ingestion date.

Why

Users ingesting files want assurance that whats stored is byte-identical to what they sent, and a way to tell whether a stored copy is stale versus a newer local file. get_checksum answers both: compare the returned digest against a local sha256sum <file>, and read ingested_at for freshness. The hash is recomputed from storage on every call (nothing cached), so it always reflects the current stored bytes.

API

get_checksum(document_id, original=false) — requires a read grant on the document.

Returns {document_id, filename, original, algorithm, checksum, file_size, ingested_at}.

• 404 if the document is missing or its bytes are unavailable (upload never completed / purged).
• original=true checksums an attached original source file instead of the parse (404 if none attached) — parity with download_document.

Design notes

• Layering: logic in documents.py::compute_checksum; the MCP tools.py fn is a thin keyword-only delegate; the @server.tool() wrapper carries the schema + LLM docstring and groups under Documents.
• Reuse: authorization + storage/key resolution reuse the existing _resolve_storage_and_key; hashing is over the Storage protocol, so it is backend-agnostic (local disk + S3/MinIO).
• Non-blocking: the fetch+hash is streamed a block at a time and offloaded via asyncio.to_thread; the DB session is never crossed into the thread.
• Auto-surfaced: no skill/catalogue edits needed — GET /mcp-tools (build_tool_catalog) introspects registered tools, so the Settings -> MCP page and the generated embedbase.zip skill bundle pick it up automatically (guarded by the existing catalogue drift test).

Decisions

• ingested_at comes from documents.created_at (the documents entry timestamp). For the presigned MCP flow this is the reservation instant, typically seconds before confirm.
• Kept the optional original=true flag for parity with download_document.

Tests / gate

6 integration tests: hash correctness, recompute-on-every-call, original, empty (0-byte) object, missing bytes (404), out-of-scope (403).

• ruff, mypy, pytest all green (970 passed, 13 pre-existing skips).
• standards-check clean; security review found no issues; code-review (high) surfaced 4 minor findings (1 fixed via the 0-byte test, 1 accepted, 2 skipped with justification).