feat(lint): add 10 Fennel, Janet and CL declaration rules

[takeokunn]

Stacked on #88 — review that first. RULE_COUNT 303 → 313. Both packages are registry-only, so no new commands and the dialect-matrix lists do not move.

package	rules
lint-fennel-janet-idiom	var-never-set (Fennel and Janet), fennel-deprecated-form, fennel-each-over-non-iterator, janet-empty-loop-body, janet-mutating-immutable-literal
lint-type-declaration	declare-not-at-head-of-body, declaim-inside-body, type-declaration-contradicts-initform, the-form-with-impossible-type, type-declaration-on-rest-parameter

Every Fennel/Janet rule is grounded in the language's own tooling rather than in taste: Fennel ships a linter plugin asserting exactly var-never-set (src/linter.fnl:80-82, pinned by its own test-var-never-set), and Janet's boot.janet:626-629 has check-empty-body calling maclintf "empty loop body" on precisely the three heads janet-empty-loop-body uses.

Six proposals dropped against a running implementation

• A late declaim ftype is not lost. SBCL flagged a later bad call identically to the early-declaim version, so the rule had nothing to detect. The only real effect is the body's own conflict dropping from WARNING to STYLE-WARNING — which SBCL already reports.
• (declare (special *x*)) with no defvar compiles silently and is the ordinary way to reference a cross-file special. Not soundly decidable at file scope.
• (optimize (safety 0)) dropped rather than shipped on alarm value — SBCL's own constraint.lisp sets it deliberately in a locally, with a comment explaining why.
• ignore-declared-variable-then-used was a true duplicate of lint-convention's ignore_declaration_conflict — same diagnosis, same category, same Fixability, literally the same worked example.
• fennel-global-set-without-global is a compile error, not a lint (compiler.fnl:676 assert-compile … "expected var <x>").
• janet-def-shadows-core would have been this batch's zero-true-positive rule — 71 core-name definitions across 241 files, the majority inside Janet's own core, which the rule cannot distinguish from shadowing it. janet-string-concat-in-loop was dropped as a duplicate of lint-performance's quadratic-accumulation after reading its body: the same loop-keyed, self-referencing-accumulator analysis, differing only in vocabulary tables. The right fix there is a dialect table on that rule, not a twin here.

One premise was refuted upward: declare-not-at-head-of-body was proposed as a warning, but SBCL errors — caught ERROR: There is no function named DECLARE. Hence Severity::Error and Malformed.

Corpus audits, and what they cost

288 .fnl and 241 .janet files from ten third-party projects; 2217 CL files with 21239 declare and 3979 declaim occurrences. Both found real false positives and both were fixed:

• var-never-set lost 10 findings to project-local macros that expand to set (tangerine.nvim's append!, jpm's setfn). Now suppressed via a macro-vocabulary read from the file's own macro bindings, guarded by symmetric controls so the suppression can't widen.
• The type rules lost all 16 of their first-pass findings: CLHS 3.2.3.1 (a top-level locally body is a top-level context), a reader conditional as the head shifting every child index, #. docstrings, and Lisp-2 operator position.

Two measurement traps worth recording. A first sweep reported zero findings because an invalid --emit json made every batch error out — a false clean. And is_unevaluated_at descends from the file root, so a linear scan there cost one pass over every top-level form per finding: 646 ms at 500 reporting definitions, now binary-searched.

Carried forward, not fixed here

• Janet's backtick long strings are unimplemented in classify_janet (packages/core/syntax/src/sexpr/reader_policy.rs), so ~21% of real Janet files get a wrong tree — 9 fail to parse outright including Janet's own src/boot/boot.janet, and 41 more parse at exit 0 with phantom nodes inside string bodies. Triple-backtick docstrings are the dominant Janet idiom, so this is not an edge case.
• The shipped ignore-declaration-conflict produces 21 apparent false positives over SBCL's sources (quoted templates, shadowing, Lisp-2 operator position, macro arguments). Corrected guards are already written in lint-type-declaration's support.rs, ready to lift across.

Both are recorded in the packages' READMEs and in the registry provenance comment so they don't get lost.

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 10 fire through the real binary, parsing the JSON findings array — var-never-set on both .fnl and .janet, confirming its two-dialect scope reaches both. Sweeping tests/fixtures/** produced exactly one new finding and it is a true positive: tests/fixtures/corpus/janet.janet:41 (var acc @[]), never reassigned — array/push mutates the array object, not the binding, so def is correct. All 12 goldens are insertions only (848 insertions, 0 deletions).

Mutation testing: Fennel/Janet 33 of 34 killed (the survivor is a documented pure-performance guard, and one mutation exposed a missing test, now added); type-declaration 8 of 8 live. One guard kills only its own unit test — traced and documented as correct-but-behaviourally-inert rather than quietly deleted.