ROFL as a linter over its own rule files, and 21 rule packs cut to their decisions

[Disentinel]

What this is

The question was whether two model-written rule files (rules/js-effects.rofl, rules/js-callgraph.rofl) could be shorter, and whether the answer would be visible structurally if ROFL itself were the linter. It was, so the instrument was built and then run over the tree.

The instrument

• scanners/rofl_lint.ts turns every rule and fact pack into facts about its clauses (what it concludes, what it reads, an alpha-normalised body, which premises are implied through every arm of another premise) and its comment blocks (what follows each, banners, dates, backticked names).
• rules/rofl-lint.rofl holds the criteria as rule bodies: orphan_block, dangling_mention/dangling_bare, section_twice/title_twice/list_repeats, twin, paired_arm, shape_twin, alias, implied, unread/unasserted_audit. Every one is a candidate list, none an [audit]: they are non-empty on an honest tree by design.
• npm run rofllint renders the report; facts/rofl-lint.rofl is the committed census and rules_rofl-lint is a golden world, so a repair shows up as a count moving on bless.

What it found and what was repaired

• Twelve orphan paragraphs in js-callgraph whose rules had moved to js-dataflow a week earlier; fn_binding cited and defined nowhere; section 6 twice and a list restarting in js-effects. All repaired.
• top_call = top_site: one rule, one name now.
• caught_here said a try with only a finally catches; repaired at the source with catches_via(try_statement, handler) and try_catches, and eff_catch_here collapsed onto it. The old finding's witness_absent became witness_in_atleast.

The comment cut

21 rule packs, every rules/js-*.rofl plus nullary, worklist, ingest, rule-shape and depends, cut to a header and one note per decision anchored to its clause. No rule text moved by a character: every golden held on every step and only the lint world moved on each bless. rules/js-* went from 6174 comment lines to 1605 against 3287 lines of rules. The instrument caught its own author six times on the way (do_while, may_be_*, resolve_*, replaced_by, rule_shape, peak_rows), which is the argument for running it on every file.

The ledger

• Five findings recorded; facts/refactoring.rofl is a new queue volume (w_queue, refactoring_queue world) holding what the lint still reports on the clause side and was deliberately left for a decision: the may_be_node/may_be_lit template, the unread relations, the implied premises, the twins in worklist (needs_reach = waits_on), depends (code_needs = generated) and ingest, and ingest's undeclared host relations.
• The service-graph design (url pack, fetch and express semantics, parts/joined abstraction, library meaning as an authored table) is recorded as five dependent items in facts/worklist.rofl and one design-principle finding.

Merge with nextver

nextver removed test/ and docs/dogfood/. The branch's edits to three test files, its mutant file for the lint (test/rofl-lint.test.ts, eight mutants) and its dogfood note went with them; the instrument's liveness now rests on the golden census alone, which is the same disposition the other 191 files got. Three settlements that cited those paths now cite the rule and the queue.

npm test: 87/87 worlds on both engines. The two DANGLING grave lines (test/version-plan.test.ts, test/publishes.test.ts) are nextver's own removed_in rows pointing at shas outside this shallow clone.

Next, per the owner: put npm run rofllint into CI.

🤖 Generated with Claude Code

https://claude.ai/code/session_011pE9mh2aKYpNhpEzgsJ6ps

---

Generated by Claude Code