Lint sweeps for STYLE042 / STYLE043 / PERF033, => line continuation, inscope after break/continue

[aleksisch]

The tree sweeps for three new lint rules (STYLE042, STYLE043, PERF033 - the rules themselves are #4213, which stacks on this branch), plus the three compiler and interpreter fixes the sweeps need and three small lint fixes their CI needs.

Commits

1. lint: LINT026 counts citations from the documents above the run's folders - check_lint_changed lints only the changed folders, so an anchor cited from a rule document one folder up (modules/REVIEW_SHADER_EMITTERS.md -> dasMetal/ARCHITECTURE.md#kernel-model-asymmetry-ledger) read as uncited whenever a dasMetal source changed. The pass now also reads the .md files above each root, as citers only.
2. dasLLAMA: the performance checklist cites setup_lcpp_ref.das by its path - LINT033: the bare name never resolved from performance/.
3. dasLLAMA: fat_start's dasllama_env require takes LINT019 like its strings_boost one - both serve only the llvm_tune half, so the nolint is stale in a build with llvm_tune.
4. parser: a => at the end of a line continues on the next - the lexer takes the line break right after => as part of the token. Every grammar rule with => requires an expression after it, so a newline there was always a syntax error; no program that compiled before changes meaning. A blank line or a comment after => still ends the statement. tree-sitter already parsed the form; its corpus and the grammar canary now pin it. The STYLE042 sweep needs it for bodies that do not fit beside their signature.
5. infer: break and continue mark early out like return - a break / continue in a nested block before a var inscope declaration left its slot uninitialized, and the finalizer deleted garbage ("can't delete locked array"). Found by the PERF033 sweep; tests/daslib/test_inscope_early_out.das pins it.
6. simulate: a while loop with a finalizer stops before its condition on return - a while body that owns a finalizer (a var inscope local) evaluated the loop condition before it looked at stopFlags, and a condition that is a real call cleared the pending return on its way out - the loop went on. read_toml spun forever on bad input in the interpreter (JIT and AOT were fine); the PERF033 sweep's inscope in daslib/toml.das exposed it. tests/language/loop_ret.das pins it.
7. STYLE042 fixes - def f(...) { return EXPR } -> def f(...) => EXPR (=> <- EXPR for a move), 936 files.
8. STYLE043 fixes - var a : T / var b : T -> var a, b : T, 257 files.
9. PERF033 fixes - an array / table declared inside a loop becomes var inscope, 223 files. Without it every iteration leaks the buffer until the heap resets.

Commits 7-9 were written by each rule's --fix and re-linted per file. They cover the code master has; the sweep of the lint code #4213 adds is in #4213.

Verification

• With the lint rules on top (lint: STYLE042 / STYLE043 / PERF033 with --fix; das-fmt --max-line-length; chunked check_lint_tree #4213), check_lint_tree is clean except for findings master already has.
• utils/lint/tests, tests/lint, ast_cursor_test, test_inscope_early_out, function_arrow_body pass on this branch alone (277/277).
• Every file outside the 62 that lint: STYLE042 / STYLE043 / PERF033 with --fix; das-fmt --max-line-length; chunked check_lint_tree #4213 touches is byte-identical to the fully tested tip.
• Full tests suite: running.