fix(lint): demote equality-arity from Error to Warning

[takeokunn]

Four PRs took this rule 1,981 → 407 by closing three structural false-positive classes: hard-quoted data (#119, #120), CLHS type positions (#121), case keys and bound variables (#123).

The 407 that remain are still false on every sample drawn — 120/120 in one adjudication, 14/14 in another. They cannot be closed the same way, and the investigation into a general mechanism is what settles the severity question.

The "unknown head is opaque" design is dead by measurement

Suppressing inside any head the engine has never seen defined silences 32,886 of 113,979 findings — 28.9% of the catalogue's output, across 186 of the 214 rules that fire. The narrow "immediate parent unknown" variant still silences 18,086 (15.9%) across 150 rules. Thirteen rules with ≥20 findings lose ≥90%: nested-function-parameter-shadows-enclosing-parameter 190/190, when-unless-implicit-nil-misused 61/61, thread-spawned-without-error-handler 35/35.

Adjudicated sample of 30 of the collateral: 23 genuine, 5 false, 2 ambiguous → ~25,000 real findings destroyed to remove 405 false ones, about 62 real per false.

And the premise is simply untrue. "A macro's arguments are not evaluated" fails for when, dolist, deftest, describe/it, macrolet, named-let, define-inline — all macros whose body arguments are evaluated. Knowing a head is a macro says nothing about which of its positions hold data.

OpacityCauseKind::UnknownHead's own doc had already reached this: "treating (print x) as opaque too would be sound as well, and would prove nothing about any real file." Common Lisp has no registry of ordinary function heads, by design.

A configured head list can't be built in scope, and wouldn't suffice

• RuleSettings values are i64 only, by explicit design ("When one is not, this grows a variant rather than a String"). The list needs packages/core/lint-engine; the config key needs config_bridge.
• A name list is the wrong granularity. Every deftransform finding is at child index 2, its lambda list — its body at index 3+ is ordinary code and yields none. define-vop needs indices 3–5, not its :generator at 6. It needs (head, position) pairs.
• Even perfect, it closes 312 of 407 (77%). The 95-finding residue is a different problem: compound type specifiers reached through or/and/not/cons/ftype combinators, plus a long tail of one-off DSL heads.

A build-blocking rule that still blocks builds on correct programs after all that is not fixed.

This is not one rule's problem

14 rules share the class, 624 findings, 7 at Severity::Error — equality-arity (312), setf-arity (77), single-arg-comparison (40), accessor-arity (21), eql-search-literal (4), eq-number-comparison (3), eql-list-comparison (2) = 459 build-blocking false positives. The shapes are identical: (setf symbol-function) is a setf function name in defknown; (>= optimizer) is a defoptimizer name; (* word) is a type specifier.

Worth fixing properly later. Until then, the honest severity for a rule measured at 407 findings and no true positives over 5,556 files is Warning.

Scale note for whoever picks this up

The 407 live in three projects — SBCL's compiler (561 of the 624), trivia (62), mgl-pax (1) — and the corpus holds five SBCL releases that content-hash dedup cannot collapse. Those 407 findings occupy 88 paths but only 36 canonical ones. The SBCL bucket is one codebase counted five times.

Counts

warning_count 255 → 256, preset-filtered 239 → 240 (the rule is untagged, so both move together). RULE_COUNT (358) and fixable_count (106) unchanged. The golden diff is 7 lines across 6 files, all the severity token.

cargo build --workspace / fmt --check / clippy --all-targets --all-features -D warnings / test --workspace all exit 0.