Lexer: keep last line of a C code file lacking a final newline

docs/GRAMMAR/ADR-records/ADR_20261111_UNTERMINATED_LAST_LINE.md

@@ -0,0 +1,27 @@
+# Keep the last line of a C code file that lacks a final newline
+
+## Status
+
+proposed
+
+## Context
+
+Inside `%{ ... %}` blocks, and in files embedded there with `%include`, the lexer (`instrument.l`, state `<ccode>`) only knew newline-terminated lines. A last line without a newline matched no rule, so flex echoed it to stdout and it was missing from the generated C.
+
+Seen with the ESS BIFROST model, where `rfc-lib.h` and `rfc-lib.c` (mcstas-radial-filter-collimator) end in `#endif // RFC_LIB_H` and `}` without a newline. McStas printed `#endif // RFC_LIB_H}` and the compiler reported `function definition is not allowed here` for everything that followed.
+
+mccode-antlr accepts such files, so they are easily produced and go unnoticed.
+
+## Decision
+
+Add one `<ccode>` rule matching a line without a newline. It can only be the longest match on the last line of a file. The rule pushes the line back with a newline appended, so it is rescanned by the existing rules (`%}`, `%include`, plain code line).
+
+## Consequences
+
+Library and code files no longer need to end with a newline. No new syntax, and nothing changes for files that do end with one.
+
+## Behaviour
+
+- An unterminated last line is kept, and gets a newline in the generated C.
+- Generated C was compared before/after for the 369 McStas example instruments: 365 identical, 2 fail to generate both before and after. 2 differ by one restored line (`Supermirror_thin_substrate`, `Test_lib_supermirror_lib`): the closing comment of `supermirror-lib.h`, which also lacks a final newline and was silently dropped until now.
+- The lexer is shared with McXtrace; only McStas was built and tested (macOS arm64).

mccode/src/instrument.l

@@ -249,6 +249,15 @@ METADATA      return TOK_METADATA;
         yylval.string = str_dup(yytext);
         return TOK_CODE_LINE;
       }
+    /* last line of a file, lacking its newline (only then can this rule be
+       the longest match): rescan it with a newline appended */
+    [^\n]+ {
+        char *line = str_dup(yytext);
+        int i = yyleng;
+        unput('\n');
+        while (i > 0) unput(line[--i]);
+        str_free(line);
+      }
     } /* end ccode */
 
  /* C character literals, e.g. 'a' or '\n', kept verbatim as a single token in