Hojun-Cho bcd948de88 wcc_ww/check: c4 #80 bare-binder forrange dealiases the iterable's type expr ...

F2a batch-4 c4. RE-PROBED AFTER c1 per spec: still live at the c3
train base with a REBUILT w6c_ww (the mechanical tichase collapse
didn't cover it — this read is AST-node-keyed, not tinfo-keyed).

REPRO (.ai/scratch/repro_f2a1_b4.ww): `untyped_lit * rangevar` over a
range-for of an alias-typed slice (`type slk = []int`) — wwstage
checker dies "asserttyped: bin" at the binop; cs accepts and runs 0.

TRACE: check.ww resolvewalk N_FORRANGE bare-binder arm — the binder's
element type comes from kind-testing the scrutinee's type expr
(N_TSLICE/N_TARRAY), but an alias-typed iterable arrives as N_TNAME:
both tests miss, the binder falls to the N_FORRANGE fallback decl,
stays untyped, and the first binop over it bails. cs twin types the
binding at scope_define (check.c N_FORRANGE) — accepts.

FIX (single site, the one the repro traces to, per grant): dealias
via the existing resolvealias(unwrapbang(it)) idiom before the kind
tests. The TUPLE-DESTRUCTURE arm carries the same unresolved tests
but is NOT in-grant (fixing the bare arm is not a no-op, so the
re-spelling clause does not apply) — FILED as task #97; currently
double-masked bounded-loud (cs louds upstream at #270-1c so the
alias-tuple-slice iterable is unconstructible on cs; ww asserttyped).

Rows (944): rangevar_alias2 (the repro, cs0/ww-reject -> 0/0 byte-id)
+ rangevar_plain_ctl (non-alias control, held throughout). 944
202/202. Corpus: five-mains ww NEUTRAL vs base on identical inputs
(checker-acceptance-only change; no alias range-for in corpus).
combined.ww regens ride along.

2026-06-06 01:04:19 +09:00

46 KiB

Raw Blame History

View Raw