perf: bound token-validation cost on adversarial input

[rmyndharis]

Summary

The CLI parses arbitrary user-supplied markdown/YAML (file or stdin). Three validation paths can be driven into super-linear CPU or unbounded recursion by a crafted DESIGN.md. Each fix is a small guard that leaves results for legitimate input unchanged.

1. Quadratic backtracking in dimension validation

parseDimensionParts (model/spec.ts) uses /^(-?\d*\.?\d+)([a-zA-Z%]+)$/; the same shape is in token-like-ignored's CSS_DIMENSION_RE. On a long all-digit string the engine backtracks every split point (clean O(n²) — measurable seconds at ~80k chars). Real CSS dimensions are a handful of characters, so a leaf value over 64 chars is rejected before the regex runs.

2. Unbounded Levenshtein cost for unknown keys

unknown-key builds a full (m+1)×(n+1) DP matrix against every schema key for each unknown top-level key, with no length guard. Since edit distance is always ≥ the length difference, a key whose length differs from a schema key by more than MAX_TYPO_DISTANCE can never be a typo — those comparisons are now skipped. Suggestions are provably unchanged.

3. Unbounded color-mix() recursion

parseCssColor recurses into each inner color of a color-mix() with no depth bound; a deeply nested value throws RangeError: Maximum call stack size exceeded, which the model's catch-all turns into a single generic error that discards every other finding for the file. A depth counter (cap 32) makes an over-deep value resolve to a normal "invalid color" finding instead, leaving the rest of the model intact.

Testing

• bun test: 285 pass, 1 skip, 0 fail (added 3 tests). The suite completes in ~0.3s, including a 100k-char dimension, a 50k-char unknown key, and a 50-deep color-mix — all previously slow/throwing, now instant/graceful.
• bun run lint (tsc --noEmit): clean.
• The new color-mix test asserts the over-deep value is rejected per-token and that a sibling valid color still resolves (i.e. no model-wide collapse).