Percent-encoded fixtures for not-found-classify (MARS-586 prerequisite, 3 pre-probe patterns)

Pluto owes percent-encoded fixtures for not-found-classify.test.ts in BOTH directions, and they are a PREREQUISITE for the MARS-586 decode, not a deliverable of it. Mars already shipped ENCODED_EVASIONS (bf31253, not-found-classify.test.ts:241-253) pinning the current bypass; pluto has zero encoded fixtures, so the gap is untested both ways -- nothing proves encoded paths currently bypass SCANNER_PROBE_RE, and nothing would catch a fix that OVER-matches. MEASURED (afe9b74): stripPath (not-found-classify.ts:426-428) splits query/hash only and never decodes, so all matching runs on the raw string. THREE pre-probe benign patterns run first on that same undecoded string -- WELLKNOWN_BENIGN_RE, ICON_PROBE_RE, NO_ROUTE_BY_DESIGN_RE (:343, probe at :357). Mars has TWO; the count is per-rail and must not be copied across. DELIVERABLE (no decode code ships): 1. raw and percent-encoded forms of the SAME path reach the SAME outcome; 2. a must-fail fixture per pre-probe pattern proving an encoded hostile path does NOT reach WELLKNOWN_BENIGN_RE, does NOT reach ICON_PROBE_RE, does NOT reach NO_ROUTE_BY_DESIGN_RE; 3. malformed encoding (e.g. /%zz, lone %) cannot throw out of the classifier. WHY THE EXISTING ORDER PINS ARE INSUFFICIENT (pmmaster): the benign-before-scanner fixtures pin the ORDER of the tests, not the STRING they run against. A decode inserted above the benign block keeps every current fixture green while changing what those three patterns match. SECOND, SHARPER FINDING, still unmeasured at runtime -- requestedPath has TWO SOURCES IN DIFFERENT ENCODING STATES: not-found-capture.ts:68 takes opts?.requestedPath ?? h.get('x-pathname'), where proxy.ts:14 stamps x-pathname from request.nextUrl.pathname (raw/percent-encoded), while the [...notFoundPath] catch-all supplies params Next has already decoded. So a single decode at one chokepoint would double-decode the other path. Verified by reading; NOT measured at runtime. The fleet design must not assume a single input string for pluto until someone measures it. Severity per pmmaster: noise gap, fails toward visible (a bypassing probe is not evading a control, it fails to be SUPPRESSED and arrives as a page). Not a security evasion. Blocked on MARS-586 adjudication for the decode itself; the fixtures are addable now.