fix(routers): detect routes registered outside main() for gin's four siblings

[nccapo]

Follow-up to #33 / #34 (now merged). Chi was not alone.

I tested three "routes outside main()" shapes against every router — a cross-package routes.Register(r) helper, a factory that builds its own router, and a method registering on a struct field:

Router	cross-pkg helper	factory (opaque signature)	struct-field method
gin	✓	✓	✓
chi (after #34)	✓	✓	✓
gorilla	✓	✗	✗
stdlib	✗	✓	✗
echo	✗	✗	✗
fiber	✗	✗	✗

Echo and fiber returned a completely empty table for the reporter's layout — the same symptom as the original issue. Gin was the only router already correct, because it walks every function and keys on types.

Two independent defects

1. The function gate chi had. Echo, fiber and gorilla walked a function only if it was main/init or its signature named the router type. A factory returning http.Handler or any is never walked.

2. Receivers recognized by variable name, not type. Echo, fiber and stdlib accepted a registration only when the receiver was a local variable assigned from echo.New() / fiber.New() / http.NewServeMux(). A router arriving as a parameter — the shape of every routes.Register(r) helper — or held in a struct field was invisible even when the function was walked.

Gorilla had only #1, stdlib only #2, echo and fiber both.

What changed

• registry.go (first commit, no behavior change): the index chi uses to find registration functions and expand their call sites is not chi-specific, so it moves behind routerIndexSpec, which asks each framework only for what differs — packages in scope, what a registration looks like, its router type. All five extractors now share one implementation.
• stdlib: receiver identified by type. The package-level http.HandleFunc case stays, now confirmed through types.Info rather than by matching the identifier text "http".
• gorilla: body-based index plus call-site expansion, so a subrouter handed across a package boundary (routes.RegisterItems(api) where api came from r.PathPrefix("/api/v1").Subrouter()) carries its prefix. isGorillaMuxType now compares the package path instead of searching the printed type for "mux.Router" — that only worked because gorilla/mux carries no version suffix, the exact trap that broke chi.
• echo + fiber: both fixes. Group variables keep their own map, since a group's prefix is what distinguishes it from the instance it was carved from. Fiber's group parameter type is the fiber.Router interface, not *fiber.Group, so both are accepted.

Middleware chains became []MiddlewareRef in gorilla, echo and fiber, so a chain assembled in one package and applied in another stays resolvable.

Safety

Widening the walk means unrelated methods sharing a registration's name come into scope — gorilla matched HandleFunc/Handle on name alone with no receiver check at all. Every extractor now confirms the receiver's type, and each new testdata module contains a decoy (bus.Handle("warm", nil), bus.Any(…), bus.All(…)) asserted not to appear.

Verification

Four new testdata modules — stdlib-multipkg, gorilla-multipkg, echo-multipkg, fiber-multipkg — each mirroring a real service layout, all six routes resolving with correct prefixes. The per-router assertions share one helper (assertRoutesOutsideMain), which also checks no route is emitted twice and none carries an unresolved caveat; the chi and stdlib tests were folded onto it rather than copied a third and fourth time.

go build, go vet and go test ./... pass. Benchmarks unmoved — extraction is a few ms against ~150ms of package loading.

🤖 Generated with Claude Code