cargo-test (full): aliased native-class import loses native methods (aliased_native_class_import, 2 tests) #8222

Activity

1. proggeramlug commented on Aug 17, 2026

   

   proggeramlug

   on Aug 17, 2026

   ContributorAuthor

   More actions

   The alias is a red herring — the unaliased form is equally broken

   aliased_matches_unaliased_for_native_class fails on its second assertion, not the first:

   assertion `left == right` failed
     left: "undefined undefined undefined\n"
    right: "function function function\n"
   

   assert_eq!(aliased, unaliased) passes — both spellings agree, and both are wrong. So this is not an aliased-import bug and not the "native base-class keyed on a literal name" weak area; chasing the alias-resolution path will find nothing.

   Minimal repro, no cargo test needed:

   import { Socket } from "net";          // note: NOT aliased
   const s: any = new Socket();
   console.log(typeof s, typeof s.connect, JSON.stringify(Object.keys(s)));

   node : object function []
   perry: object undefined []
   

   Object.keys(s) is [] even when it works — Socket methods are never own properties. They resolve through the property-get path keyed on the binding's ("net","Socket") native-class type (perry-codegen/src/expr/property_get_names.rs → is_net_socket_method_name). There is deliberately no "Socket" arm in lower_builtin_new. So the failure is that the binding loses its native-class type, not that construction builds the wrong object.

   Bisected

   git bisect run over the verified window, judging on connect: function vs undefined, with a coherent rebuild (perry + perry-runtime-static + perry-stdlib-static + perry-ext-net in ONE cargo invocation) at every step:

   690882d6d  #8018  GOOD  connect: function
   a11093a16  #8042  skip  (build failed)
   9c1a672ce  #8043  skip  (build failed)
   3381e1b89  #8033  BAD   connect: undefined
   

   Prime suspect: #8033 fix(codegen): require runtime evidence for local binding types. Its own diff says "Declared TypeScript types are not runtime evidence." That is exactly this mechanism — the native-class binding type is inferred, not runtime-proven, so it stopped being trusted and the property-get arm went dark.

   Caveat: #8042 (async-hooks ALS) and #8043 (nextjs lazy path modules) are skipped, not eliminated — their builds failed so the bisect could not judge them. They are implausible by subject matter, but unproven.

   Likely repair, given #8033 shipped scripts/local_binding_type_audit.py + a 549-entry allowlist: admit native-class construction as legitimate evidence rather than revert. new <NativeClass>() from a resolved native-module import is runtime evidence of the receiver's identity.

   Watch for a wider blast radius

   If a native-class binding type is no longer trusted, net.Socket is unlikely to be the only casualty — the same property-get shape backs other native classes. Worth checking net.Server, tls.Socket, http.ClientRequest in the same run rather than fixing Socket alone.

   Verified window

   The last observed-passing run is 08-12 c4b2c1c8e (aliased_net_socket_keeps_native_methods ... ok). The 08-13 run was cancelled and never ran the test, so "no FAILED line" there does not mean passing — grep for a positive ... ok, not for the absence of FAILED.
2. proggeramlug

   closed this as completedin #8278on Aug 17, 2026