An audit of every comment against the code it describes turned up fifteen inaccuracies, all fixed here; the comments were otherwise sound, including the JIT's register maps and instruction encodings. Stale counts: the compiler and test-suite module docs still said "both engines" from when only the Pike VM and backtracker existed, now "all four"; and the lazy DFA's state cap was described as "five hundred" in three places when default_max_states is 512. Misplaced doc comments, where a paragraph had drifted onto the wrong declaration: the fused-repeat description sat on ExitIfSame (the empty-loop guard) rather than RepOp; an SSE register-budget note sat on the BitsCache dedup table (the same note already lives correctly on first_const_xmm); and orphaned paragraphs sat above maxSimdRanges in the x64 backend and above asciiRunSet in the JIT runtime. Moved or removed as appropriate. Behavioral drift: the DFA's stepped_empty flag is set when every thread dies including the freshly-seeded one, not just a pre-seed thread; the x64 ASCII run scanner does thirty-two bytes under AVX2, not only sixteen; the oracle's corpus-mutate mode compares two ways (leftmost and every match), not the three the generated sweep does, corrected here and in CLAUDE.md; the oracle module doc showed the -Dpcre2 override flags as the primary invocation rather than the from-source default; and a test comment said `.*?` reports "all of ab" where it reports `a` then `b` in turn. Comment-only; the full suite and zig fmt --check pass. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PTEzA4fBK8SqdvRvF1aicm