benchmarks.md's table listed block_encode and block_decode, which do not exist in src/bench.zig and never appear in the doc's own sample output. Moved them to the "no upstream counterpart yet" section with the reason they are covered transitively, and stated that the table is exactly what `just bench` prints so a count mismatch reads as doc rot. Also defined "admission-critical", which the verify/unchecked rationale leans on without ever saying what it means. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>