feat(lint): add 4 Hy and LFE lint rules

[takeokunn]

Stacked on #92. RULE_COUNT 316 → 320, registry-only, no new commands.

This completes dialect coverage — every one of the ten dialects now has at least one rule written for it specifically. Before this session's batches, 13 of 286 rules ran on anything but Common Lisp.

rule	dialect	severity	notes
hy-mutable-default-argument	Hy	Error
hy-identity-comparison-with-literal	Hy	Warning
hy-bare-except	Hy	Warning
lfe-catch-swallows-exit	LFE	Warning	pedantic

Premises were run, not argued

Both interpreters were obtainable — Hy 1.3.1 and LFE 2.2.0 / Erlang 27.3.4.15.

The Hy mutable default is the mirror image of a rule this project already refuted for Common Lisp. CL re-evaluates an &optional init form per call, so the shared-mutable-default defect does not exist there; Python evaluates a default once, at definition time, so the identical shape is a real bug in Hy. Confirmed: three calls return [1], [1 1], [1 1 1]; {} and #{} behave identically, None does not.

Also confirmed: CPython emits SyntaxWarning: "is" with 'int' literal through Hy; (except []) catches KeyboardInterrupt and SystemExit while (except [e Exception]) catches neither; and (catch (exit 'boom)), (tuple 'EXIT 'boom) and (catch (tuple 'EXIT 'boom)) are the identical term, so an LFE caller cannot tell failure from success.

Refuted and dropped: (if 'false 'a) returns false in LFE rather than raising, so if-without-else is not a defect; ! to a dead pid returns the message with no error and isn't statically detectable; case/receive without a catch-all is "let it crash" and would be mass false positives.

Also stale in the original brief: Hy 1.3.1 rejects #@ decorators and with-decorator (the current form is (defn [decorator] name …)), and mangling is foo? → hyx_fooXquestion_markX, not _is_.

The corpus audit changed three of five candidates

2825 .hy across 284 repos, 2701 .lfe across 195 — nobody involved wrote any of it.

rule	candidates	findings	repos
hy-mutable-default-argument	884	1	1
hy-identity-comparison-with-literal	712	2	2
hy-bare-except	607	32	11
lfe-catch-swallows-exit	321	146	9

• hy-mutable-class-attribute was killed — 33 findings over 251 candidates, 0 genuine defects. The premise is true and I measured it ((is a.items b.items) really is True), but in real Hy a mutable class attribute is a declaration: __slots__, __all__, Django ModelAdmin fields, Textual BINDINGS, Gym metadata. A test pins the decision so it isn't reintroduced.
• hy-mutable-default-argument was narrowed from "mutable default" to "mutable default the body mutates", taking it from 88 findings to 1. The dropped ones were hand-checked — [notes [0 5 7]], [breaks ["amen.wav" …]] are read-only lookup tables. The survivor is genuine: [known-ids #{}] with (.add known-ids id) in a body whose own docstring shows the author expected a fresh set per call.
• lfe-catch-swallows-exit was tagged pedantic rather than killed. 146 findings, but only 9 of ~143 repos and two of the three heaviest are LFE's own implementation.

Both dialect gates were proven on the corpus, not just in unit tests: 38 catch candidates in Hy files → 0 findings; 5 is candidates in LFE files → 0.

The pedantic tag makes the preset arithmetic non-obvious

The divisor moves too, so none of these is "previous + 4":

• recommended rule_count 301 → 304 = 320 − 16 (was 316 − 15). Rises by 3.
• recommended warnings 217 → 219: suite warnings rise by 3 to 235, but one of those 3 is itself pedantic, so 235 − 16 = 219. Rises by 2.
• fixable count stays 103 — this batch is the first to actually exercise that assertion's claim, adding a pedantic rule while the number holds, because the new pedantic rule is ReportOnly.

Confirmed independently in the regenerated goldens: broad.json.golden gained exactly three rule rows, not four.

A latent contract bug this exposed, flagged not fixed

feature_dependency_contract scans manifests as whole text rather than parsing them, so the package's comment explaining a removed dev-dependency read as two declared feature→feature edges. Rewording it to name the packages by directory fixed the build — no allowlist entry was added, since the edges don't exist and that would have weakened the contract.

But an intermediate wording containing the bare marker paredit-feature- made the scanner emit an empty-string dependency name and fail differently. declared_feature_dependencies (tests/cli/feature_dependency_contract.rs:53) cannot tolerate the marker appearing without a name after it. Pre-existing; nothing in the tree triggered it before.

Reader limitations recorded, not worked around silently

In the package README, with measured counts: Hy's interpolated f-strings unimplemented (390 of 2825 files fail to parse), #! shebangs not stripped (393 files, exit 0 with junk atoms), Hy bracket strings #[[…]] parsed as code (33 files — the package defends against this explicitly, since a rule could otherwise fire on text that isn't code, and a test pins it), Hy's ~ not a ReaderPrefix so QuoteState.quasi never counts down for Hy, and LFE #B(…)/#M(…) orphaned from their list in 243 files, inflating arity at exit 0. Plus two more found in passing: LFE |quoted atoms| split at whitespace (95 files) and #\(/#\) restructuring the tree.

Same class as the Janet backtick defect fixed in #91, with the same disposition — a reader repair belongs in core/syntax.

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 four fire through the real binary, parsing the JSON findings array, and the pedantic gate is confirmed in both directions: lfe-catch-swallows-exit fires under --preset all and is correctly absent under recommended. Zero findings sweeping tests/fixtures/**, where the two relevant fixtures are real negative controls rather than empty — corpus/hy.hy has an immutable default and a typed except, both correctly unreported.