fix(lint): stop 38 more autofixes deleting a reader prefix

[takeokunn]

Completes #122. The same defect — RuleFix::single(view.span, …) replaces a region starting at the form's own reader prefix while the replacement text is built from the inner form — is in 38 further rules. The remedy is view.content_span, whose extent coincides with span on unprefixed code.

Census

5,699 unique files (16,115 raw, SHA-256 deduped): 42 rules dropping a prefix across 2,650 of 17,751 replacements. Two are #122's; two — sharp-quoted-lambda, redundant-quote — exist to delete a prefix, so content_span would defeat them and they are excluded.

Worst rates: multiple-value-list-of-values 100%, redundant-prog1 89.5%, redundant-the 80%, if-not 65.8%, single-operand-boolean 42.7%, subseq-zero 41.7%, constant-if-test 41.3%.

Result

	baseline	after
files SBCL read before, not after	29	2
prefix-dropping fixes (of the 38)	706	0

The 2 are a strict subset of the 29 — zero regressions.

Three rules did not take the plain swap

• empty-let builds its fix as a sub-span inside its domain.
• nested-progn operates on a child node — took fix(lint): stop the progn fixes deleting a reader prefix #122's redundant_body_progn treatment, a prefix requirement on the child.
• constant-when-test has two fix sites.

append-nil showed zero drops in this corpus but carries the identical span and is named in the residual read-break list. Changed anyway; the absence of corpus evidence is recorded rather than glossed.

Anti-over-suppression, keyed on the finding span

The naive check is useless here: the span change moves every replacement offset by construction, so counting by replacement would report the migration itself as a behaviour change. Keyed on the finding span instead:

• Span change alone: fixes on unprefixed ordinary code 14,080 → 14,080. Not one lost. Exactly one fix changed corpus-wide — the nested-progn exemption.
• Splice guard: removed exactly 243 fixes, all ,@. 0 non-splice removed, 0 splices survived, 0 added, 1,799 legitimate fixes retained.

Controls: every changed rule with tests has a "still fires inside a quasiquote template" case asserting the rewritten source — e.g. `(when c (or ,x)) → `(when c ,x) — plus symmetric negatives (a plain ,x operand is still one operand; an unprefixed child under a prefixed parent still fires), so neither guard can widen unnoticed.

Oracle

SBCL's own reader with a custom readtable recording byte offsets of '/`/,/,@/#', cross-validated against this parser over 80,321 regions in 4,581 files — up from 1,194 in #122's audit after adding package-error recovery that lifted readable coverage from 40% to 80%.

0 positional disagreements, 0 byte-identity mismatches. Six apparent mismatches are all ,. (unquote-dot), where positions agree but the parser labels it Unquote rather than UnquoteSplicing — real, and separate.

The 2 remaining read-breaks are a different bug

Dotted-pair notation read as a call, reported not fixed:

(append . nil)                  →  (copy-list .)
(cons . (list sequence t))      →  (list . sequence t)

The parser exposes . as an ordinary child atom, so any rule counting operands is wrong there. It also reaches redundant-body-progn. Fixing 2 of an unknown number across 106 fixable rules would sprawl this PR — it deserves its own audit.

Mutation

12 mutations, all killed — including M7 (widen the splice guard to any reader prefix) and M9 (widen nested-progn's guard to the enclosing form), the two over-suppression mutations, both caught by the quasiquote-template controls. One mutation was initially reported INVALID because cargo fmt had reflowed its anchor; re-anchored and killed. Compile failure is matched on error[E/could not compile and reported INVALID, never as a kill.

Scope

50 files, all under packages/feature/lint-*. fix list still reports rule_count: 106; no golden and no pinned count moved — so no repo fixture contained one of these corruptions.

The lint-control-flow/src/support.rs conflict with #122 was two copies of the same helper with different doc comments; bodies verified byte-identical (42 lines) before keeping the merged one.

cargo build --workspace / fmt --check / clippy --all-targets --all-features -D warnings / test --workspace / test --test cli (3,085 passed) all exit 0.