[YPARK]

data-beans as a set of tools an agent can use, and a view a person can watch it from. The minor bump is for the new subcommands and flags.

data-beans mcp is a stdio MCP server on legume-mcp 0.1. Every tool runs a child data-beans, so stdout stays the protocol, a cancel kills the child, and log lines come back as progress.

• Read tools: info, list_contents, names (regex, paged), statistics (quantiles and top entries), values (at most 100 entries).
• Write tools:
  • qc: a dry run, then writing the kept cells.
  • plot: returned as a PNG image.
  • import (records its source), subset (names, a pattern, or index ranges), subsample, split, convert, merge (common_rows).
• No tool replaces a file. An output that is an input, or anything else named after the output beside it, is refused. squeeze is not offered.
• Records, after Noble (2009): every write goes into the day's replayable results/<date>/runall.sh and results/INDEX.tsv; unnamed outputs go in results/<date>/; import notes provenance in data/README.md.

data-beans mcp watch (legume-mcp's tui view) lets a person follow a live server:

• its calls, logs, results and notebook;
• s stops a call, n leaves the agent a note (it arrives with the next result), o opens a written backend in the statistics explorer;
• --watch-only drops steering.

New CLI:

• qc: senna's and pinto's cell QC, with the same --qc-* flags, as a summary or a filtered backend. The mito pattern defaults to (?i)^MT-.
• plot qc and plot stat: figures as .png, .svg or .pdf, through legume-plot.
• merge-backend --common-rows: keeps the rows every input has, in the first input's order.
• subset-rows and subset-columns take index ranges (-i 0-19612,20000, cut at the matrix's end; repeats refused) and --pattern (a case-insensitive regex). Selections conflict, and are resolved before anything is copied.

Shared qc_lib (senna and pinto get these too):

• --qc-mad-on-genes=false and --qc-mad-on-counts=false now parse; they used to be stuck on.
• New --qc-mad-on-mito.
• A band with no spread (MAD 0) no longer drops every cell that differs from the median.
• A test fails on any CLI flag that defaults to on but can't be turned off; --allow-prefix was one, and is fixed.

Tested in Rust (cargo test):

• the MCP tools end to end over stdio, including a sweep of every tool's refusals;
• cancellation;
• the records;
• following a live server through the watch pane and leaving a note.

Also tried for real with Claude Code as the agent on the scVelo pancreas data (3,696 cells × 39,226 rows), through every tool. A replay of its runall.sh in a fresh copy of the project reproduced all 15 backends and 3 figures exactly.

cargo fmt and cargo clippy --all-targets are clean. The branch went through /simplify and /code-review at each step.