Epic: Evidence — prove the ReDoS-safety and speed claims (P1)

[lbliii]

Part of the stack-wide evidence program (pounce #227 pattern: name the claim, name the incumbent, publish the artifact, cite it from the README, gate regressions). Rosettes is the priority case: its claims currently outrun its proof. The README asserts "O(n) guaranteed / Zero ReDoS / ~12ms tokenize on 10k lines" with no proof doc, no CI benchmark job, and no committed artifact — while the raw material (test_redos.py, lexer property invariants, 55-language golden fixtures) already exists and is genuinely strong.

The claim: ReDoS-safe, O(n) tokenization across 55 languages; competitive with Pygments on wall-clock; 1.5–2× parallel speedup on 3.14t.

Tasks

• Proof doc for the O(n)/Zero-ReDoS claim: adversarial corpus + the existing property invariants + test_redos.py written up with method — and the scope caveat stated plainly (55 languages ≠ Pygments' ~600; the claim holds within coverage)
• Pygments comparison artifact: same corpus, wall-clock + token-parity differential diff (patitas' oracle pattern); committed baseline; README table replaces the flat ~12ms numbers with cited, caveated ones (patitas' honest voice is the model)
• Wire benchmarks into CI with a regression gate (kida's committed-baseline pattern)
• Claims ledger (public-claims.json) citing both artifacts

Acceptance

Every performance/safety claim in the README cites a committed artifact or proof doc; a benchmark regression fails CI.