Repair regressions from the fleet sweep: art blocks, Args bodies, comment lines, term lead-ins, ignore at level 0, docs-only README variant (closes #27)

[thorwhalen]

Closes #27. Part of #16 (WP6 fleet sweep).

The principle, now written down

epythet/normalizer.py states it and every rule enforces it: rewrite only what is unambiguous; otherwise report. What a rule leaves alone, epythet validate reports (DR014, DR002, DR031), so nothing is silently dropped.

The eight defects

#	Defect	Root cause	Fix	Fixture
1	blank lines inserted inside ASCII-art diagrams	blank_lines_between_blocks fired on the varying indentation of drawing lines	new ART line context (is_art_line, _art_runs); every rule treats it as code	ascii_art_untouched
2	Args: entry name: + indented description became name::	literal_block_after_colon knew nothing about Google section bodies	google_section_bodies(); the rule skips section bodies and lone term: lines	google_args_bare_name_untouched
3	README snippet claimed the repo "ships tooling" when it only publishes agent-readable docs	one snippet for every case	agentic-readme-section-docs-only snippet, chosen by section_snippet_for()	test_docs_only_project_gets_the_shorter_section
4	# comment lines in prose became rubrics (33 in one i2 run)	markdown_headings_to_rubrics accepted any # line with alphanumeric text	heading shape required (capitalised title, no code characters, no trailing :/., no TODO: tag); a lone # needs blank lines on both sides; never next to another # line or after code	comment_lines_untouched, markdown_h1_heading
5	:return: nested into a preceding Note: body	google_one_liners folded field-list lines as continuation	only wrapped prose is folded; a field list, bullet or heading ends the section	note_one_liner_then_field_list
6	specifying: + indented paragraph became a :: literal block	no prose-vs-code test on the indented block	the block must read as code (no line with four or more plain words); a lead-in over a paragraph is a definition list and DR014 reports it	term_then_indented_prose_untouched
7	docstring-style module template tripped DR014	template put Main entry points: directly over an indented block	template is now blank line + bullet list (what the swept repos already use)	DR014.py: bad_entry_points_over_indented_block, good_entry_points_blank_line_then_list
8	validate -i tests/ not honoured at level 0/0.5	the level-0 linters (ruff, pydoclint) ran without the ignore list; parsing was fine in both argument orders	is_ignored() is the one predicate; lint findings are filtered through it; -i a -i b now accumulates for validate and repair	test_ignore_applies_to_the_linters_at_level_0, test_ignore_parses_the_same_in_every_argument_order

Also found while running the dry-runs

• repair stripped an author's blank line before the closing quotes and counted it as a rewrite; the count of trailing blank lines is now preserved (test_trailing_blank_lines_before_the_closing_quotes_are_kept).
• Blank lines were inserted between wrapped Args: entries; consecutive definition-list terms need none (google_args_wrapped_entries_untouched).
• bare_headers_to_rubrics is no longer source-safe: napoleon renders a bare Examples: header as that rubric already, so the rewrite churned the source for no change on the page (now in UNSAFE_RULES with the reason; still runs at build time).

Verification

• 548 tests green, module doctests green.
• repair --dry-run over fresh read-only copies of dol and i2 (after their sweeps): 4 rewrites remain, all correct RST fixes (a blank line after a bullet list before prose; before a field list after an indented paragraph; before a nested list after can:; # Signature Calculus as the first line of a module docstring to a rubric). None of the eight patterns.
• Independent adversarial review (separate agent, brief: refute the change) returned FIX-THEN-MERGE with three should-fix items, all folded in with fixtures: a lone Usage: / Output: lead-in over code keeps its literal block (_looks_like_code, usage_lead_in_literal); a lone # heading needs a blank line before it, not after (markdown_h1_glued_below); a doctest, bullet or field line ends a drawing run, and a bare >>> prompt is never art (arrows_then_doctest, bullets_with_arrows, bare_prompts_untouched). Its master-vs-branch comparison of build-time output over every docstring in dol, i2 and meshed (2215 docstrings, stale build copies included) now differs in 25, each one an intended fix.
• Two of its nits are left as they were, on purpose: is_ignored() matches tokens against the absolute path (pre-existing in file discovery; a relative match needs the project dir threaded through iter_python_files), and google_section_bodies() is a separate pass rules opt into rather than part of line_contexts (a LineInfo refactor is the right next step if a fourth rule needs it).

Judgment call to know about

A # Title line that meets every heading condition (first line of a module docstring, blank line after, capitalised words) still becomes a rubric. If rubrics from # lines are unwanted altogether, the switch is one line: add markdown_headings_to_rubrics to UNSAFE_RULES.