Ripscope’s own scanner reads every file as cscope does

To give cscope’s answers, Ripscope reads C the way cscope’s scanner does. That scanner never refuses a file. Code it cannot follow costs accuracy where that code is, and the rest of the file is still indexed.

Decision

Ripscope has one scanner, and it is Ripscope’s own code. With every Correction off, it reads C as the reference cscope does, measured file by file against cscope’s own database. Corrections plug in at named decision points (ADR 0002).

The scanner is written from observed behaviour, not translated from cscope’s source. Its behaviour description is in Ripscope’s own words and backed by probe fixtures of the reference’s output. Its rules, names, comments and module structure are its own. Keyword and delimiter tables and mark bytes are compatibility facts, recorded as observed. Where cscope’s result depends on which of several look-ahead patterns matches longest, a small layer evaluates each group by longest match with rule priority.

The scanner never refuses a file. Linear-time caps bound its work without dropping a file. Where cscope hangs or its behaviour is undefined, as with NUL bytes, and where it truncates long names, Ripscope defines its own behaviour. These cases are listed in differences from cscope.

Consequences

• With Corrections off, a difference from cscope is a scanner bug.
• Every readable file contributes navigation facts, so Ripscope sometimes gives a wrong answer where a stricter reader would give none. Corrections improve those readings where the compass measures them.
• The scanner is covered by Ripscope’s licence alone.

Considered options

• Port cscope’s flex scanner. The port would be a derivative of cscope’s code under its licence, and would need an emulator for flex’s start conditions, yyless, unput and jumps between rules.
• A grammar-based parser such as tree-sitter-c. Many real files can be parsed only by evaluating combinations of preprocessor branches, which grows combinatorially.
• Refuse a file the scanner cannot read exactly. The file would lose every definition, call and include, and the line protocol cannot show that a result is missing.