Collector: dispatch on parsed statement type, not first token

[rmichaelthomas]

Summary

_collect_checkable_statements decided whether a line was checkable by testing whether its first token was a verb. starting "<date>" and until "<date>" are TokenType.CONNECTIVE, and inherited is TokenType.OPERATOR — both consumed statement-initial, before the verb, by _parse_one_operation (parser.py). A first-token-VERB guard therefore silently dropped every temporally-prefixed or inherited-prefixed statement: no error, no warning, just absent from the collected list — with check_source still reporting encodable=True, concealing the under-reporting.

Found by the Halverson re-run (liminate-dev PR #92): four statements across three documents excluded from every check, combined checked=17 where it should have been 21. inherited shares the identical defect and was never exercised by any corpus — fixed here too, not left for a second finding.

Changes

• Inverts the collector: tokenize → reorder → parse every line (as before), then dispatch on the resulting AST node's type — the same pattern check_agreement already uses downstream via _VERB_NAMES and _iter_statements.
• DefineNode registers into predicate_names before statements.append, preserving the forward-declaration order exactly.
• run._collect_deontic_statements, check_agreement, the encoder, and the seven checks are untouched — this changes collection only.
• Removed the now-unused _CHECK_DEONTIC_VERBS constant and TokenType import.

D3 — a guard against the class

Added a hand-written coverage table (test_collector_coverage_matches_parser_accepted_shapes, 42 parametrized cases) asserting the collector checks exactly one statement of the expected type for every deontic verb, define, and value-form remember, under every prefix combination the parser accepts (bare, starting, until, starting+until, inherited, starting+inherited, full stack). Deliberately not derived from parser.py internals — a specification of intended coverage, not a mirror of the implementation, so a future grammar addition without a matching row here is a gap the test can still expose once someone notices, not one it silently absorbs.

Test plan

• Regression test: a starting-prefixed forbid returns checked=1, not 0
• Same for until, starting+until, inherited, and the full canonical stack
• A temporally-prefixed statement participates in findings (unless_swallows_rule fires correctly)
• A define on a temporally-prefixed line still registers in predicate_names for later lines
• checked count unchanged for a source with no prefixes — no double-counting, no newly-swept-in non-deontic lines
• The C3 corpus re-verified: combined checked=21 (was 17) with the fix applied
• Full suite: pytest tests/ -q → 1837 passed (1787 prior + 50 new)

🤖 Generated with Claude Code