docs(args-contract): correct two audit findings that keyword-rosetta had already settled

[squid-protocol]

Follow-up to #2773 / #2786, and a correction to my own work in it.

The 46-language audit in docs/args_rule_contract.md reported four contract violations and filed four issues. Two of them were already triaged in keyword-rosetta's deviation_ledger.json — the corpus's audit trail of exactly these questions — and I did not check it before filing. Both are corrected here.

haskell — inside the contract, not outside it

haskell-caf-bindings-count-as-functions (engine-semantic, validated 2026-09-04):

Top-level value bindings (CAFs like region = unsafePerformIO ...) count func_start/args like any function binding — morphologically correct for haskell, where top-level values ARE nullary functions.

That reading is better than the one I filed. In Haskell a top-level value genuinely is a nullary function, and the engine is consistent about it: func_start counts the same binding, and _args_arrow_count_groups correctly derives arity 0 from the arrow-less type. The narrowing I proposed (require a top-level ->) would have made args and func_start disagree on the same binding — worse than the thing it fixed. #2785 closed as not-planned.

m4 — the headline number is intended morphology; a narrower thing survives

m4-parameters-are-use-sites (intended-morphology, validated 2026-09-04):

m4 macros have no parameter list [...] $1 inside the body IS the parameter, so the rule counts one arg per USE, not per declaration [...] equalising it would mean writing macros that never reference their own argument.

So the +38% file-level figure is not a defect, and m4 owes no re-bless. What the ledger does not cover is avg_func_args, which is an arity: m4_define(probe_test, [AT_SETUP($1) AT_CHECK($1)]) is one parameter referenced twice, and reads 2, because m4 lacks the _args_findall_max_groups helper shell uses for the identical morphology. #2784 rescoped and retitled to that.

Unchanged

groovy (#2782) and apex (#2783) stand — genuine call-site counting from a member arm whose return-type prefix can match zero times. keyword-rosetta now carries them as args-call-site-counting-apex-groovy, split out of ts-callparen-args when #2786 retired that entry's typescript and objective-c arms (corpus PR pairing this one).

What changed in the doc

• The haskell and m4 audit rows, and corollary 3 — which now says what the ledger established: at file level a positional reference count is the language's own morphology; it is avg_func_args where counting references rather than distinct positions is wrong.
• The findings section, which now records both corrections rather than quietly dropping them.
• A new entry in "Notes for the next rule": check the ledger before calling a language's args a defect. grep '\"signal\": \"args\"' deviation_ledger.json is a 30-second check that would have caught both of these, and it outranks any amount of fresh measurement.

Docs only — no engine or test changes, no golden-master impact.

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab