Sub-ms latency for a pipeline this dense is impressive. I’m particularly interested in the recursive decoder. How does Glyph handle the 'Late Canonicalization' problem?

For example, if a payload is encoded in a way that looks like high-entropy 'noise' to a detector (bypassing the printable-ratio gates), but is then decoded by a specific plugin 'sink' after the guardrail has already fired?

Also, regarding the homoglyph fold: does it catch cases like the fraction slash (⁄) being used in place of a solidus (/) for path traversal? In some of my recent research into AI orchestration frameworks, those homoglyphs often bypass standard NFKC normalization but still resolve to valid paths once they hit the OS file system.