docs: why the CLI is one file, and that its split is a date rather than a rule

[esamoun]

src/darkvessel/cli.py is 2042 lines and 21 subcommands. That is the first thing a reader judges about the package, and nothing in the repository said anything about it. This adds one docs/decisions.md entry that measures the file rather than leaving the reader to assume either way, and names what its shape does not justify.

What it records, all of it measured on this branch:

• 2042 lines against 10,040 in the rest of the package — 17% of the source
• 54 functions; the longest is main at 178 lines and every line of it is the argparse table
• the longest handler is _archive_run at 94, then _retrieve 84, _crops 78
• 40 modules imported across 54 import lines: wide and shallow, the depth elsewhere
• 7 of the 11 *_request_from builders in the file are imported directly by the tests; docs/training-runbook.md imports an eighth

What it admits rather than glosses over:

• three functions compute instead of delegating (_span, _exclusion_over, _clustering_over)
• eleven builders live in cli.py and three live beside the code they feed, and the only thing separating them is the date they were written: everything up to 2026-08-28 went into cli.py, everything from 2026-08-29 did not, and nobody went back. None of the three has a caller inside its own module, so the placement changes nothing that runs.

The decision: the eleven stay. The move is 339 lines across eleven modules plus a changed import path in thirteen test files, and no test can fail on reverting it — which is a fair description of its value. The two conditions that would change it are written down.

Documentation only. No code, no config, no run journal, no figure touched: the diff is 56 added lines in docs/decisions.md and nothing else.

make lint clean, make test 572 passed / 1 skipped — unchanged.