Noir::ZigCalleeExtractor

A call expression: an optional @ (builtin) + identifier, then any number of .identifier accessors, immediately followed by (. The lookbehind stops the match from starting in the middle of an identifier or chain, so foo.bar( yields foo.bar once, not also bar.

(?:pub )? (modifiers)* fn name ( — captures the function name. The extern "C" calling-convention string is already blanked by strip_non_code, so only the keyword spacing has to be tolerated.

Zig keywords that are followed by ( in normal code (if (…), while (…), switch (…), catch (…)) and would otherwise be reported as callees. fn/return/try etc. never precede a call paren directly but are kept here for clarity.

Receiver roots whose calls are pure noise for endpoint review context. std.* (std.debug.print, std.mem.eql, std.fmt.*) appears in nearly every handler and drowns the meaningful callees.

test { … } / test "name" { … } block opener. Route registrations inside a test block are unit-test fixtures — and, in a framework's own source vendored as a loose file (modules/httpz.zig, modules/router.zig), its self-tests — never runtime endpoints.

A .zig file that belongs to a vendored copy of a framework checked into the source tree (zig-pkg/zap/…, src/deps/tokamak/…) rather than to the application. Such trees ship the framework's own tests and examples (zap/src/tests/test_auth.zig's .path = "/test", tokamak/example/'s @"GET /:name"), whose route literals would otherwise surface as phantom app endpoints. The standard fetched-dependency cache (.zig-cache) is already pruned upstream; this matches the manual-vendoring layouts — a vendor directory immediately followed by a framework package directory.

Find the index of the delimiter matching the opener at open_index. Returns nil on imbalance. Operates on the already-stripped char array so braces inside strings/comments are gone.

Convenience map keyed by simple function name. When two functions share a name (e.g. several zap endpoint structs each defining get) the first wins; callers that need struct scoping should filter function_table by offset instead.

Same as function_table but reuses an already-stripped char array from prepare / strip_non_code, so callers that already paid for a strip don't re-walk the file.

Line counter for Regex::MatchData#begin positions. Avoids the text.chars allocation every endpoint emission used to pay for. #begin is a CHAR index, so convert it to a byte offset before the byte walk — otherwise a multi-byte char before the match (e.g. inside a preserved string literal) makes the char index under-shoot the true byte position and the reported line number comes out too small (regression class of #2297). \n is single-byte ASCII, so the byte walk itself stays correct and allocation-free relative to Array(Char).

Like strip_non_code but keeps the contents of double-quoted string literals — route paths live inside "…", so the framework analyzers scan this form to read the URL while still ignoring routes that sit in a comment, a \\ multiline doc-string, or a char literal.

Byte ranges of test blocks, brace-matched on the string-blanked source so a {/} inside a literal can't throw the matching off. in_test_block? then lets an analyzer drop a route whose registration sits inside one.

Takes the scan-base-relative path (Analyzer#base_relative_path), never the absolute one — the vendoring layout is a property of the project, not of the directory the checkout happens to sit in.