codegen(deno-esm): Deno externs missing from codegen lookup → ReferenceError at runtime (endsWith / stripSuffix / numToFixed2) #470

Summary

Three pub extern fn declarations in stdlib/Deno.affine (endsWith, stripSuffix, numToFixed2) have no entry in lib/codegen_deno.ml's lookup table. The codegen happily emits call sites for them, but never defines them in the compiled output. At runtime: ReferenceError: endsWith is not defined.

Repro

A minimal port (hand-trimmed from the live standards/scripts/check-ts-allowlist.affine port that hit this):

use Deno::{ endsWith };

pub fn main() -> Int {
  if endsWith("foo.ts", ".ts") { 0 } else { 1 }
}

$ affinescript check minimal.affine
Type checking passed

$ affinescript compile --deno-esm -o /tmp/min.deno.js minimal.affine
Compiled minimal.affine -> /tmp/min.deno.js (Deno-ESM)

$ deno run /tmp/min.deno.js
error: Uncaught (in promise) ReferenceError: endsWith is not defined
    at main (file:///tmp/min.deno.js:N:M)

Root cause

lib/codegen_deno.ml has a comment at ~line 541-542:

// not externs — endsWith/stripSuffix/pathJoin/etc. are NOT here:
// they are real AffineScript built on `ends_with`/`substring`/`++`.

The intent is clear: these externs are supposed to be backed by real AS bodies in stdlib/string.affine (e.g. fn ends_with). But the Deno-ESM backend doesn't link stdlib bodies for extern fn declarations — only for pub fn definitions actually imported via use.

So we get a worst-of-both-worlds:

• The codegen lookup table (b "..." lines) covers pathJoin, walkRecursive, regexMatch, etc. — those get inline __as_* JS shims.
• Real stdlib pub fns (e.g. string::split) get their bodies emitted when imported.
• But pub extern fn declarations whose "real implementation" lives elsewhere fall in a gap — neither pathway covers them.

endsWith, stripSuffix, numToFixed2 are the three that fall in the gap today (per comm -23 of the two name lists).

Workaround (used in standards#310)

Drop the use Deno::{ endsWith }, inline a string_sub-backed helper:

fn ends_with(s: String, suffix: String) -> Bool {
  let slen = len(s);
  let sfxlen = len(suffix);
  if sfxlen > slen { false }
  else { string_sub(s, slen - sfxlen, sfxlen) == suffix }
}

This is exactly the defensive pattern the merged seed in standards#283 used; my rewrite removed it on the assumption the extern would link. It didn't.

Proposed fix (pick one)

1. Add the missing lookup entries to codegen_deno.ml — pure mechanical fix, mirrors what's done for pathJoin:

   b "endsWith"     (fun a -> Printf.sprintf "String(%s).endsWith(%s)" (arg 0 a) (arg 1 a));
   b "stripSuffix"  (fun a -> Printf.sprintf "((s,x) => s.endsWith(x) ? s.slice(0, -x.length) : s)(%s, %s)" (arg 0 a) (arg 1 a));
   b "numToFixed2"  (fun a -> Printf.sprintf "Number(%s).toFixed(2)" (arg 0 a));

   Cheap; ships with the next compiler bump; no stdlib linkage required.

2. Compile-in the real AS bodies when a use Deno::{ endsWith } is encountered. More general; would also unblock similar gaps in other stdlib modules. Larger change.

Option 1 is the right STEP. Option 2 is the right LONG-TERM. Recommend doing option 1 now and tracking option 2 as a separate stdlib-linkage issue.

Related

• standards#283 — seed PR (used defensive inline helpers; merged but didn't compile)
• standards#310 — compile + runtime fix that re-introduced the workaround
• standards#311 — workflow swap that ships the compiled output

Acceptance

• codegen_deno.ml gains lookup entries for endsWith, stripSuffix, numToFixed2
• Minimal repro above compiles and runs (deno run returns 0)
• A test in tests/codegen-deno/ covers each (probably extends deno_scripting.affine)
• Follow-up issue filed for general extern→stdlib-body linkage (option 2)