feat(lint): add 8 Clojure and Scheme lint rules

[takeokunn]

RULE_COUNT 295 → 303. Two new packages, eight standalone inspect commands. All eight are scoped away from Common Lisp, taking the dialect-specific report list 6 → 14.

package	rules
lint-clojure-idiom (CLOJURE_ONLY, report-only)	with-open-returns-lazy-seq, def-inside-function-body, single-key-nested-path, apply-with-literal-collection
lint-scheme-idiom (all fixable)	scheme-begin-single-form, scheme-let-star-independent-bindings, scheme-memq-assq-literal-key (Scheme only), scheme-named-let-never-recurs

Both packages were audited against real third-party code, and both lost a rule to it

Corpora: 1714 .clj/.cljc from 12 cloned projects (clojure/clojure, clj-kondo, leiningen, reitit, malli, ring, core.async, next-jdbc, clj-http, timbre, tools.cli, compojure) and 295 files / 3.0 MB of Guile 3.0.11's stdlib. Candidate counts are reported per rule so a clean sweep can't be a false-clean — e.g. with-open-returns-lazy-seq swept 162 real with-open scopes for its zero.

scheme-eq-on-number-or-char-literal dropped — 462 candidates, 15 findings, zero true positives. All 15 were characters or small fixnums, inside the range Racket normatively guarantees eq? for. After removing the guaranteed cases nothing syntactically detectable remained. Replaced by scheme-memq-assq-literal-key, which R7RS §6.4 settles outright ((memq 101 '(100 101 102)) is unspecified where memv is defined) and which found a real latent portability bug in Guile's own punify.scm:65.

count-compared-to-zero dropped on stronger evidence — its recommended repair throws. count accepts anything Counted; seq/empty? need seqability, and core.async's FixedBuffer/DroppingBuffer/SlidingBuffer implement Counted and nothing else, so 12 findings recommended code that does not run. clojure.core:6433 then refutes the premise outright: empty? is defined as (if (counted? coll) (zero? (count coll)) (not (seq coll))) — the rule was flagging the body of the function it recommended. Narrowed to the sound subset, 0 of 55 findings survived over 11091 candidates.

Two further Clojure defects invisible to the package's own tests, both fixed and both checked for over-suppression: a case test-constant list ((case tag (def defonce ...) ...)) read as code, and (comment ...) bodies read as code. Two Scheme false positives likewise fixed — a named let inside a syntax-rules template, and (begin ,form) at quasiquote depth 0. Re-audit: 10 findings, 10 true positives, 0 false positives.

def-inside-function-body has an independent oracle: clj-kondo's :inline-def fixture expects rows 5, 8, 10, 12, 14, 16, 20; we report 5, 8, 20 — a strict subset, so zero false positives and four documented false negatives.

A silent-corruption bug, proven with the real applier

scheme-named-let-never-recurs computed a fix_span covering a comment in the (let <gap> loop ...) gap. Verified by building a probe driving the real dispatcher and apply_byte_span_edits — the same applier workflow.rs:1186 calls — writing real files. With the guard mutated out, all three comment forms lose bytes on disk and the reparse write guard passes every time: the corrupted-Lisp-still-reparses failure mode this repo has been bitten by before. tests/fix_apply.rs pins all three, including the #; datum-comment case that was missing.

Two things the wiring exposed that the rules alone would not have

• fixable_rules_match_the_fix_engine drove only Common Lisp and Emacs Lisp fixtures, so the four rules declared Fixable emitted no fix at all — a --fix that silently does nothing. Added a Scheme fixture rather than relaxing the assertion; fix plan now reports fix_count: 4.
• a_dialect_specific_report_is_supported_for_exactly_one_dialect had hardened an accident into a rule. Every scoped rule until now happened to name exactly one dialect; three of these name Scheme and Racket. Renamed to ..._for_a_proper_subset_of_dialects asserting (1..DIALECTS.len()) — the lower bound still catches an under-supported report, the upper bound a general one. Deriving the expectation from lint_rule_dialect_scope would have been circular, since support_status reads that same declaration.

Cost

Mutation testing: Scheme 37/37 killed, zero survivors; Clojure 41 mutations, 36 killed, 5 survivors all documented cost guards. It found a real gap — deleting defmethod from def-inside-function-body's Heads left the entire suite green, because the domain tests walk the tree themselves and never touch the head index. The rule would have silently stopped seeing every defmethod. Now pinned for all four rules.

Doubling ratios are flat (Scheme ~1.07–1.12×, Clojure per-invocation 0.80–1.24×) — nothing superlinear.

Verification

cargo build --workspace, cargo test --workspace, cargo test --test cli (3083 passed), cargo fmt --all --check, cargo clippy --all-targets --all-features -- -D warnings — all exit 0.

All 8 fire: one fixture each through the real binary, via both inspect lint --preset all and the standalone inspect <rule> command, one finding apiece. The .rkt column is the load-bearing one — it confirms the three two-dialect scopes actually reach Racket and that memq-assq is correctly silent there. Zero new findings sweeping tests/fixtures/**. All 12 goldens are insertions only.

Known gap, not introduced here

#:ns{...} namespaced map literals fail to parse, which silently drops whole files from any Clojure lint run (clj-kondo/corpus/nested_namespaced_maps.clj). A paredit-core-syntax reader gap, untouched by this PR.