MilkDrop fidelity: read preset files the way MilkDrop reads them

[CaYatur]

What

With fidelity on, preset files are now read the way MilkDrop 2 reads them. The rules come from Nullsoft's code (jecassis/foo_vis_milk2 5b44cea): _GetLineByName, the GetFast readers, ReadCode and the Import functions, which are the same in the D3D9 code. With fidelity off, the old parser stays.

MilkDrop builds an index of the file's lines first. A line's name runs up to the first =, space or line break. It then looks keys up in its own order, trying the line after the previous read before scanning from the top. As a result:

• Case: keys are case-sensitive, so PSVERSION_comp and fwarpAnimSpeed are never read.
• Line shape:
  • an indented line gets an empty name and is never read;
  • key value is read;
  • key = value is not read.
• Duplicates: a key written twice takes the next line or the first occurrence, not the last one.
• Numbers:
  • integer keys are read with %d, so textured=0.05 is off (it was on for us);
  • float keys take the leading number, so .975; reads as 0.975.
• Code:
  • numbered code ends at the first missing number;
  • equation lines lose // and \\ comments and are glued with nothing between them.

Which presets change. Readings differ in something the engine uses (readingsDiffer) for 32 corpus presets and none of ours. The causes are duplicated keys, fractions in integer keys, key case, a per-frame name written in the header, shapes past MilkDrop 2's four, a \\ comment and a gap in the numbering.

Where the reader is used.

• The Preset records which reader built it (readAcc).
• Every Preset built in src states its reading rule explicitly, and a test enforces this.
• Turning fidelity over rebuilds the running preset only when the two readings differ.
• The stage choice reads the file the same way.

Mash-ups read their donors the same way. A preset gives a part only if MilkDrop reads that part from it, and version lines come from the donor's MilkDrop reading. Because the same recipe can now produce different version lines, the rule version goes to 3.

Left out on purpose (listed in the ROADMAP)

• Byte rules. MilkDrop reads bytes, while the engine gets decoded text. Its two byte rules (0xFF ends the file; a value over 251 characters splits into a second index line) can't be kept exactly, and neither affects a corpus file.
• Float storage. Values stay double, where MilkDrop stores floats; the difference is below 1e-7.
• Compile failures. A block that fails to compile is still recovered statement by statement. Our parser and MilkDrop's disagree on what counts as an error; for example, three presets call _aboeq, one of MilkDrop's internal functions that our compiler lacks. Dropping whole blocks would therefore drop some that MilkDrop runs.

Measured

• Against main (last 10 of 60 frames):
  • control presets are identical with fidelity on and off;
  • of the 32 presets that read differently, 15 change with fidelity on (12 by more than 1%, up to 32%), and all 32 are identical with it off;
  • the 900-preset sample keeps its classes, preset for preset.
• Mash-ups on the whole corpus:
  • the part test agrees with the engine for every preset and part (0.04 ms per test);
  • a mash-up of one preset reads as that preset for all 10,332;
  • 5,000 random mash-ups keep every part and stage of their donors.

Tests

• 10 new tests; 23 of 23 mutations are caught.
• npm test: 2197 of 2197 pass, and each commit passes on its own. The changed files also pass under Linux emulation.
• npm run smoke: PASS, with the userData checksums unchanged.

Refs #580
Refs #579
Refs #560