TypeScript/JavaScript literal receivers bind a confident false edge — `"x".replace()` hits an unrelated user function (#59)

[joyful-ii-V-I]

What is wrong

In every TypeScript and JavaScript codebase ripwire indexes, a built-in method call can link to an
unrelated user function that happens to share its name. "a-b".replace(…), [3, 1, 2].map(…) and
/x/.test(s) all do it. The graph's ambiguity gauge stays at zero, so the wrong edge reads as
confident, and an agent is sent to code that never runs. This is issue #59.

An earlier fix vetoed built-in method names wholesale and was backed out, because it deleted true
edges: on webpack, --callers=stringify went from 361 to 0. The safe first step is to decide by the
receiver's type, not its name — starting with receivers whose type the syntax already proves,
which is to say literals. Everything else keeps today's behaviour.

The evidence

test/fieldnarrowcheck.sh carries a KNOWN GAP block on a freshly generated fixture, rewritten
minimally from #59's own reproduction:

• Five arms pin today's wrong behaviour — a string, a template, an array, a regex, and a chain
  starting from a string literal. Each binds an unrelated function in src/unrelated.ts, with
  graph_ambiguous="0". They pass today. Flipping them is the finish line.
• Two controls pin true edges any fix must keep: a typed user-object receiver r.replace() binding
  Rewriter.replace, and "x".shout() binding the repo's own String.prototype.shout.
• A presence guard keeps the block from passing vacuously.

Confirm the arms pass today with bash test/fieldnarrowcheck.sh on a plain dev build.

Size

Medium. The code is small: a receiver classification at extraction (an append-only enum, with a
version bump and its mirror) and one resolver arm. Most of the work is measurement — establishing
the full population of edges the change removes on a real corpus, not a sample. webpack is the
historical corpus for this one.

Where to start

prompts/help-wanted/ts-literal-receivers.md is a self-contained prompt for a coding agent. It
carries why this matters, the file pointers, the rejected attempt and why it failed, the
reproductions, the design space and constraints, the acceptance criteria, the known traps — every
one of them measured on main while the kit was written — and what the fix PR's description should
contain. Like every prompt in prompts/, it ends by writing a plan and stopping — a maintainer
agrees the plan before any code is written.

Nothing else has to merge first; the gap is on main today.

Comment here to claim it.