{$...} directives are processed inside (* ... *) comments

Type: bug (Pascal frontend — lexer/comment handling) — Track P (edits the shared lexer.inc, so A's gate + no-concurrent-edit rule)
Status: done
Opened: 2026-07-11, second blocker of the New-ZenGL Pascal ladder ([[feature-game-library-candidate-suite]] slice C).

Symptom

A {$...} directive inside a (* ... *) block comment is evaluated instead of being skipped as comment text:

program tcmt;
(* dead block
  {$if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L}
  junk
  {$endif}
*)
begin
end.

pascal26:0: error: conditional directive: unexpected character ()

FPC treats a { } (and any {$...}) inside a (* *) comment as plain comment content — one comment type may appear inside the other. PXX's lexer processes the brace directive anyway, so commented-out junk (here: machine-translated C preprocessor lines with && in a Khronos-generated ZenGL unit, zgl_gltypeconst.pas:612-660) kills the compile.

Where hit

library_candidates/zengl/Zengl_SRC/srcGL/zgl_gltypeconst.pas — a large (* ... *)-commented block of half-translated C #if lines. Blocks every ZenGL unit that pulls zgl_gltypeconst.
Synapse (2026-07-11 re-probe at v201): external/synapse/jedi.inc has a (* ... *) documentation block (lines 48-699) containing example directives, including {$IF Declared(RTLVersion) and (RTLVersion >= 14.2)} (line 501) — PXX evaluates it, chokes on 14.2 ("unexpected character", reported at synautil.pas:458 through the include splice), and EVERY Synapse unit dies at {$I jedi.inc}. This regressed the Synapse compile-check from its previous wall ("too many array constant elements") to line 52 — earlier probes (v83) got past jedi.inc, so either the comment/directive handling changed since or the doc block did. Fixing this unblocks the whole [[feature-synapse-compile-check]] ladder at once.

Acceptance

{, }, and {$...} inside (* ... *) are comment text (and vice versa: (*/*) inside { } already behave per the nested-comments setting).
The ZenGL repro above compiles; self-host byte-identical (mind the [[project_nested_comment_brace_selfhost_landmine]] class of hazards: comment handling changes can desync the 2-step self-host — reseed, don't bisect).
A lexer test covers directive-in-()-comment and brace-in-()-comment.

Resolution (2026-07-12, opus-night)

TWO distinct bugs were behind the symptom; both fixed:

Include-expansion pass processed directives inside comments/strings (compiler/elfwriter.inc): the pre-lex {$include} expansion scanned raw text for {$ with no awareness of (* *) comments, // line comments, or string literals (it already skipped { }). ZenGL's zgl_gltypeconst and the tcmt repro died here. The scan now skips all three (mirroring the lexer's comment rules; the main lexer's (* *) scanner was already correct).
{$MODE DELPHI} leaked NestedComments across units (compiler/parser.inc ParseUsesUnit): Synapse's jedi.inc sets MODE DELPHI (NestedComments off) and every lib/rtl unit lexed AFTER it broke on nested-brace doc comments ("unexpected character" at the first orphaned }). Fix: per-unit scoping like the existing CaseSensitiveMode precedent — caller state saved/restored AND the child starts from the FPC default (True), since FPC lexes each unit under its own mode, not the includer's.

Also defused two nested-brace doc comments in platform_backend.pas that only lexed by grace of NestedComments=on. Result: tcmt repro compiles; ZenGL zgl_gltypeconst advances to its real next wall (uses X — no X11 unit); Synapse synautil now parses ALL the way into semantics (undefined variable DayOfWeek — an RTL surface gap, Track B). Gate: 2-step self-host byte-identical + testmgr FULL 1131/1131 GREEN.

Log

2026-07-12 — resolved, commit ee65eed0.