Eof (standard input) not recognized

• Type: feature (compiler builtin / RTL — Track A)
• Status: done
• Owner: Track A
• Opened: 2026-06-21
• Relation: next examples/chess blocker after the bare-function-name result-var fix (c8ccc7b). Chess now parses to its REPL.

Problem

examples/chess/chess.pas:895 uses bare Eof to test standard input:

while running do
begin
  write('chess> ');
  if Eof then Break;     { pascal26:895: error: undefined variable (Eof) }
  readln(line);
  ...

PXX does not recognize Eof (no-arg, standard input). readln/writeln/read are compiler keyword-builtins (tkReadln, …); Eof has no such token and no RTL symbol, so it resolves to nothing.

Direction

• Simplest: a no-arg Eof: Boolean builtin/RTL routine that reports end-of-input on stdin (and ideally Eof(f) over a textfile, consistent with the textfile unit). Sibling of the read/readln builtins.
• Track question: an unrecognized identifier in expression position points at a missing builtin (Track A). But it is textfile/stdin I/O, which is Track B's lane (see the textfile unit work). Decide ownership before starting; if it is modelled as a textfile method on the implicit Input, it is Track B; if a bare-stdin intrinsic, Track A.

Acceptance

• if Eof then ... over standard input compiles and behaves (true at end of piped input).
• examples/chess advances past chess.pas:895 into its REPL loop.

Resolution (2026-06-21, Track A)

Implemented as a no-arg special builtin (call id 210, like Length), x86-64 codegen — the same support level as readln (both x86-64-only; cross errors gracefully with "builtin/special call not yet supported", no crash).

• Parser (ParseFactor): bare Eof / Eof() (no file arg) lowers to an AN_CALL with id -210, Boolean result. Eof(f) (a ( + argument) still falls through to the textfile Eof(var f: Text) routine; idx<0 lets a user variable named Eof win.
• Codegen (EmitEof, ir_codegen.inc): result in rax. Not-eof if the line buffer holds unparsed content (pos < len) or a pushed-back byte is held; else peek one byte from fd 0 — read ≤ 0 is EOF, a real byte is stashed in BSS_PEEK_BYTE / BSS_PEEK_VALID and reported not-eof.
• EmitReadLine consumes the pushed-back byte before reading, so the byte Eof peeked is not lost (verified: x\ny with no trailing newline reads y in full).

LANDMINE (cost a debug cycle): the BSS-allocation edit that added BSS_PEEK_VALID/BYTE accidentally dropped the BSS_LINE_LEN line, so len aliased another global (nonzero) → pos<len always true → Eof skipped its read and readln segfaulted. Restored the allocation. When inserting BSS slots, do not disturb the existing ones.

• Gate: make test green; default + --threadsafe self-host byte-identical. Tests: test/test_eof_stdin.pas (piped, incl. no-trailing-newline + empty input).
• chess: advances past chess.pas:895; next blocker is eng.Free — built-in TObject has no Free method (see bug-method-call-free-tobject, backlog).

Log

• 2026-06-21 — filed (self-discovered as the next chess blocker after c8ccc7b).
• 2026-06-21 — implemented (special builtin 210 + pushback). Track A.