node:http radar (#800) under PERRY_NO_AUTO_OPTIMIZE strips ext-http-server; symbol stubs mask real failures #2156

Summary

scripts/node_core_subset.py runs every test with PERRY_NO_AUTO_OPTIMIZE=1 (set unconditionally at script line ~262). That disables build_optimized_libs, which is what triggers the well-known-bindings flip that pulls perry-ext-http-server (and friends) onto the link line.

Without the flip, symbols like js_node_http_create_server / js_node_http_server_listen / js_node_http_server_address_json fall to the symbol-stub-generator in compile.rs, which emits a stub that returns undefined. The compile succeeds, the test runs, and http.createServer().listen(...) fails at runtime with undefined.listen.

This is partly why #2129's radar attributes 15 failures to the server-listen cluster — many of those programs would compile and run normally with --auto-optimize enabled.

Options

1. Drop PERRY_NO_AUTO_OPTIMIZE=1 from the runner. Trade-off: per-test rebuilds are 10–30s on a cold cache, so the http subset would go from ~5 min to ~1 hr without the flag. Realistic if scoped to specific APIs.
2. Pre-warm target/perry-auto-<hash> for each API's required feature set before the loop, then drop the flag for that API only.
3. Pre-link the ext crates into libperry_stdlib.a so the flag becomes safe even without auto-optimize (would inflate the prebuilt stdlib size considerably).

Probably (2) is the right trade-off — it preserves the radar's wall-time goal while not lying about which features are reachable.

Related

• node:http: 57 method-surface gaps (Agent/ClientRequest/IncomingMessage/Server.listen) #2129 (this skews the radar's bucket counts for http/https/http2)
• compat: Node core test subset comparison job #800 (radar tracker)
• node:http server-method dispatch bypasses native_table NR_PTR / class_filter under typed-feedback path #2153 (the typed-feedback dispatch issue is independent — these are two different "undefined.listen" causes)