Enumerate the promo surface by output-diff, not compile-success

Blocker for [[bug-nilpy-int-promotion-decided-statically-so-computed-overflow-wraps]], whose cost question is now decided ([[decide-nilpy-int-promotion-costs-10x-on-ordinary-loops]]) but which cannot land until this is done.

Why it exists

That ticket's survey concluded "only hex, bin and oct break". It was wrong, and the way it was wrong is the point: it checked whether the program compiled. Re-run comparing OUTPUT against CPython, four more shapes were silently wrong — str(i+1) and round(i+1) printed the slot address, "ab" * (i+1) printed nothing, and [0] * (i+1) raised TypeError.

A compile-success survey cannot see any of those, and this frontend's whole risk model is about silent wrong values. So the survey has to be redone properly.

The shape every one of them has

A promotable int reaching either:

• a hand-built call site — FindProc returns ONE proc and never consults overloads ([[project_findproc_by_name_ignores_overloads]]), so the promo lands on a parameter that neither narrows nor boxes it and the SLOT ADDRESS is used as a value; or
• a static-type predicate — promo is deliberately not reported by TypeIsOrdinal / TypeIsPyNumeric, so a promo operand classifies as "unknown" and a path is chosen for it that assumes it is not a number.

Both are consequences of the rvalue model (an rvalue is the SLOT ADDRESS), and both are invisible to a compile check.

What to do

1. Apply the promotion patches (devdocs/progress/patches/int-promotion-option1-*).
2. Run a WIDE corpus — the .npy test suite is the obvious start, plus a generated probe over every builtin and operator against every operand shape — and diff stdout against CPython, not exit status.
3. For each divergence, add the arm: narrow at the call boundary, box to a variant, or teach the predicate about promo.
4. Re-run until the corpus is clean, THEN land the patches.

Note step 3 is not always a one-liner: adding promo to IRPyOperandKind was tried for the [0] * n case and turned the TypeError into a garbage number, so the list-repeat path has more than one thing in the way.

Gate

The whole .npy suite diffing identical to CPython with the promotion patches applied, plus the bug ticket's own table, plus a benchmark line recording the measured cost so the 10x figure stays honest as check-elision lands.

Log

• 2026-08-04 — resolved, commit b3ccf2f61.