repair regressions found by the fleet sweep: blank lines inside ASCII-art blocks, Google Args `name:` + indented body turned into `name::` literal; agentic-readme snippet overstates tooling

[thorwhalen]

Discovered from #16 (WP6) while sweeping i2mint/meshed (i2mint/meshed#78). Three defects to fix with the next epythet touch, each with a fixture in the ledger corpus:

1. epythet repair / repair_package inserts blank lines inside ASCII-art diagrams in docstrings (the doctest/list "blank line before" rule fires on lines that start with > or - inside a literal drawing). Guard: never rewrite inside a block that is already a literal block, a fence, or a run of lines with box-drawing/ASCII-art characters; when in doubt report, do not rewrite.
2. Google Args: entries written as name: followed by an indented description on the next line get rewritten into name:: (an RST literal block marker). The one-liner-section rule must only apply to section headers (Returns:, Yields:, Raises:, Examples: ...), never to argument names inside a section body.
3. The agentic-readme default snippet says the repo "ships tooling for coding agents" even when the only agentic aspects are the agent-readable docs (llms.txt, <pkg>.md). The snippet needs two variants: one for repos that ship skills/agents/instruction files, one for repos that only publish agent-readable documentation (shorter, no "ships tooling" claim).

Sweep-session details: the local crowsnest ledger for sweep-meshed. Until fixed, sweep sessions must review repair --dry-run diffs for these two patterns before --write.

[thorwhalen]

More from the dol sweep (i2mint/dol#100; details in the local sweep-dol ledger):

4. repair_package turns # comment lines in prose (not inside a doctest or fence) into heading rubrics; the ## Heading rule must require a space after #, at least one word, a blank line after, and must not fire on a line that looks like a code comment (following an indented code line or inside a literal block).
5. repair nests :return: (and other RST field-list lines) into a preceding Note: block because the Google-section indentation rule treats the field list as the section body. Field lists end a Google section.
6. Term: prose (a line ending in a colon followed by an indented paragraph, i.e. an ordinary RST definition list or an intentional lead-in) is rewritten into a literal block. The one-liner-section rule must only fire on the known section names.
7. The epythet-docstring-style skill's module-docstring template ("Main entry points:" followed by a list) trips rule DR014 in validate; make the template and the rule agree (blank line + list, or a rubric).
8. epythet validate -i tests/ (ignore) is not honoured at Level 0.5; the ignore list must apply to file discovery for every level.

Fix all eight with fixtures in the ledger corpus (bad + control examples), then run repair --dry-run over the dol, i2 and meshed copies in the baseline scratch dir and confirm the diff contains none of these patterns.

[thorwhalen]

Detail for item 8 from the i2 sweep: epythet validate . -i tests/ scrap/ (ignore option after the positional) is not applied at levels 0/0.5, while epythet validate -i tests/ scrap/ -- . is. Looks like the repeatable option swallows or drops values depending on position; fix the CLI parsing (cw Sequence[str] option) and add a usage test for both orders.

[thorwhalen]

Evidence from the i2 sweep (commit 6bf444c5 on i2mint/i2, before hand-fixes): repair --write inserted 33 .. rubric:: lines. Sources: Google Example:/Examples: headers followed by an unindented doctest (9); # comment lines in prose, including commented-out doctest outputs like # True, # open, # x=1, y=(2, 3), z=4 (16); wrapped prose paragraphs where each line became its own rubric (ContextFanout, 5 lines). It also turned specifying: + indented paragraph into a :: literal block (extract_arguments). Use these as fixtures: the rubric rule must never fire on consecutive prose lines or on #-prefixed lines that are not Markdown headings (require # + capitalised words + blank line after, and never inside or right after a doctest).