The shape is a three-step pipeline. Step one is a security predicate: a regex, a validate_file(), a filter_var($..., FILTER_VALIDATE_IP), a startsWith('/api/'), an includes('..'), an ACL lookup against a route grammar, a signed-prefix comparison. It inspects an input against a forbidden form (path traversal segments, private address ranges, dangerous HTML tags, out-of-scope paths, forbidden URL schemes). Step two is a canonicalization: url.PathUnescape, path.Clean, str_replace, PHP's urldecode, Go's net/netip.ParseAddr, Win32 LCMapStringEx, Node's fs.realpathSync, the kernel's getaddrinfo, marked.parse, Array.prototype.join, JavaScript's decodeURIComponent, Perl's ExifTool tag resolver, NTFS ADS separator handling. Step three is the dangerous sink that the canonicalized form reaches: locate_template, fs.rm, HTTPConnectionPool.connect, innerHTML, CreateFileW, fs.openat with truncate, a Kerberos ticket issuance, a proxied upstream fetch.
The three are usually written by different authors. Often they live in different files, sometimes in different processes or different machines. The predicate's author believes the forbidden form will reach their check unaltered. The canonicalizer's author is doing something the subsystem needs for its own correctness, not for security: a URL library lowercases the scheme because that is what URI normalization means; a kernel performs case-folded Unicode comparison because that is what Unicode equality means; an archive extractor resolves NTFS ADS separators because that is what NTFS semantics require; a template engine resolves CSS nesting because that is what the specification defines. Each transform is defensible in isolation. Composed with a gate that runs before the transform, they produce the forbidden form on the attacker's behalf.
The failure repeats across stacks because canonicalization is pervasive and gates are local. Any system that routes, parses, decodes, resolves, or normalizes before acting has at least one place where the canonical form differs from the wire form. Any gate that compares strings instead of structured semantic representations has a candidate gap. The catalog exhibits trace the same shape through HTTP routers (Atlassian's webresource Router inventing a :: to / rewrite that Tomcat does not speak, CVE-2026-21589; UniFi's auth gate reading x-original-uri while Nginx dispatched against the normalized $uri, CVE-2026-34910; Alfresco Share's regex denylist on ..;/ that Tomcat collapsed to ../, CVE-2026-26336), filesystem helpers (SillyTavern's /api/extensions/delete running the truthy check before sanitize-filename collapsed . to the empty string, CVE-2026-44650; WinRAR's file_name_check deleting trailing dots and spaces to manufacture the path traversal its component-start rule was supposed to catch, CVE-2025-6218), string manipulation (Visual Composer's str_replace('theme:', '') producing the ../../../ that validate_file had just approved, CVE-2026-12227; SiYuan's escapeAriaLabel running before a decodeURIComponent call that manufactured the angle brackets the encoder was supposed to prevent, CVE-2026-44588), SSRF validation (MLflow's webhook validator resolving a hostname via getaddrinfo only for requests.post to resolve it again against DNS-rebind-flipped authoritative records, CVE-2026-64849; Statamic's filter_var IPv6 range tables not classifying IPv4-mapped prefixes before PHP 8.3, CVE-2026-45660), reverse-proxy middlewares (s3-proxy's auth layer reading r.URL.RequestURI() while the handler read chi.URLParam + url.PathUnescape, CVE-2026-42882; Nitro's radix3 router keeping %2F opaque while the proxied upstream decoded, CVE-2026-44373; Traefik's StripPrefix running JoinPath on the stripped bytes to produce /admin from /api../admin, CVE-2026-48020), and off-URL grammars entirely (Microsoft's CompareStringW using NORM_IGNORECASE while lsass's SPN lookup used 0x31403's wider flag set to normalize enclosed-alphanumerics back to local-machine identity, CVE-2026-26128). The transform categories vary. The structural defect does not.
The pattern's distinguishing feature against Gate Narrower Than Sink is the presence of an intermediate transform. Gate-narrower-than-sink pairs two matchers describing the same bytes through different grammars: a brace counter against a CSS compiler, a startswith('SELECT') against a Trino SQL parser, a \s+ regex against an \s* regex. The matchers may see identical input; they simply recognize different sets. Gate-before-canonicalize requires a canonicalization step that produces the forbidden form only on the sink's side. The gap is a function call, not a grammar difference.
The structural fix is to canonicalize before checking. Apply the sink's transforms to the input, then run the predicate against the output. Spring's uriDecode exhibit at CVE-2025-41242 shows what happens when canonicalization is partial: the author truncated sixteen-bit Java chars into eight-bit bytes before the check and Jetty 12 finished the canonicalization afterwards, manufacturing the IIS escape .%u002e that Spring's pattern set did not cover. Partial canonicalization moves the gap; it does not close it. The costlier fix is a bouncer: refuse any input where the two parses would diverge, which s3-proxy, Nitro, Open WebUI, and UniFi all chose when unifying the parsers was architecturally infeasible. The gate and the sink stay disagreeing about strings nobody is allowed to send. The class signature is the same; only the implementation substrate rotates.