nx_lang_lint_ext.nx

nx_lang_lint_ext.nx -- TWO landmine classes nx_lang_lint does not yet see, both paid for in build cycles on 2026-07-25/26. This EXTENDS the existing linter's vocabulary; it does not replace it, and it deliberately re-implements none of L1-L4. L5 EMPTYESC -- an EMPTY LITERAL IN DISGUISE. L2 catches a literal "" by looking for two adjacent quote bytes. It does not catch "\x00", which is four source characters and still a zero-length string, so it hits the SAME nx_cc defect (a zero-length literal returns a pointer to the NEXT literal in the pool). That is precisely the shape that segfaulted a gate: a sentinel spelled "\x00" silently became some other string. A detector that catches only the obvious spelling of a defect leaves the subtle spelling free to keep landing. L6 UNDEFCONST -- an ALL-CAPS identifier that is USED but never DECLARED anywhere in the file or in what it imports. nx_cc compiles these silently (seq1012, sev 9) and resolves them to whatever is to hand: first a wild pointer, then -- after unrelated code moved the frame -- the caller's output buffer, so the program ran and wrote WRONG DATA. A typo that does not fail the build is worse than one that does. HONEST SCOPE, stated rather than implied: L6 covers CONSTANT-STYLE names only ([A-Z][A-Z0-9_]+). A mistyped lowercase function or variable is NOT covered yet. That is a real limit, and it is written here so nobody reads a clean run as more assurance than it is. Constant-style is where the failure actually happened, and it is the subset that can be detected without resolving the whole language. Declarations are gathered from the target AND its transitive imports, because an identifier arriving from an import is not undefined -- a checker that cannot see through imports reports noise, and a checker that reports noise gets switched off. license_tier: ORIGINAL expect_exit: 0