[Ljunqiang]

Fixes #922

Root cause

BuildableChecker::check_binary_operands compared numeric families, not concrete kinds:

arm	old mismatch term	effect
< <= > >=	lk.is_numeric() != rk.is_numeric()	int vs float64 both numeric → never mismatched
+	same	x + f accepted
- * /	kind membership {Unknown, Int, Float} only	x - f, x * 2.5 accepted

All six issue repros checked clean (0 diagnostics), then go build refused each with invalid operation: ... (mismatched types int and float64) — GUARANTEES.md Row 13 (BUILDABLE).

Fix

New Collector::numeric_adapt_conflict(lk, rk, lhs, rhs) — applied as an else if in the three arms — rejects cross-kind numeric pairs unless Go's only implicit numeric conversion bridges them:

spelling	verdict	why
1 + 2.5, a + b (both untyped consts)	accept	untyped constant expression
f + 1, 1 + f, f + k (untyped const k = 1)	accept	int const → float64 adaptation
x + 1, x < 10, s + "b", s < "c"	accept	same-kind / string families
x + f, f + x, x - f, x * f, x / f	reject	two typed operands
x < f, f <= x, x > f, x >= f	reject	two typed operands
x + 2.5, 2.5 - x, x < 2.5	reject	untyped float const against typed int
f + k with const k int = 1	reject	annotation makes k typed

Define.untyped_const (new field, from the already-recorded LocalDecl.ty) distinguishes fn-local const k = 1 from const k int = 1 — the annotation existed but was never consulted (tightens #1112's fail-open on this site; assignment sites keep #1112's rule until a dedupe).

Validation (all on this branch @ 47eb793)

• hy-check/tests/binop_operand_types_issue_922.rs 24/24 (15 reject faces, 9 accept controls)
• cargo test --workspace green except pre-existing local-CRLF spec_t3_strings (reproduces on clean upstream/main)
• corpus pin binop-mismatched-numeric-reject.hy — EXPECTED-VERDICT: REJECT BUILDABLE + HYB-E1002; census 126 → 127 on all 7 surfaces (parse-all assertion included)
• harness 127/127, tripwire not fired
• negative flip: baseline (upstream/main buildable.rs) accepts the pin → harness 126/127 with the pin FAILing; restoring the fix → 127/127
• differential 40/40 accept (build+run) + 87/87 reject

Credit / coordination

• Harvest credit: stale companion rickymulya PR fix(hy-check): reject mismatched operand types in binary arithmetic and comparisons #923 (bot:stale-base + bot:conflict) proposed the same rule family with literal-only helpers — this take supersedes it with untyped-const-ident support (f + k) and the fn-local annotation check.
• Census race: fix(emit): flush slot loans between operand eval and return, and on ? exits (#1398) #1428, fix(check): atomic values are copy-locked — reject every by-value copy (#1399) #1430, Fixes #1112: typed int values no longer satisfy float64 targets (untyped-constant clause across all 20 value positions) #1447 also bump 126 → 127 — whichever lands last rebases the count (128+), same pattern as the fix(hy-check): reject assignment arity mismatches, single-value bindings, and parallel type conflicts (fixes #905) #906/[Soundness] hy-check: parallel assignments and multi-target declarations bypass count and type conflict checks (HYB-E1002) #905 sibling take.
• Dedupe: numeric_adapt_conflict carries a NOTE against Fixes #1112: typed int values no longer satisfy float64 targets (untyped-constant clause across all 20 value positions) #1447's is_untyped_constant (near-identical walk) — the kinds_conflict / PR fix(hy-check): close the call-arity and return-statement accept-side bypasses (#577 C9/C10) #620 duplicate-helper precedent; whoever lands second collapses the pair.

[limitlessprojects]

Queue-reset sweep: this repo now runs a rolling conveyor, and open PRs
outside the active review window are being closed as not planned to keep the
queue honest — a stale queue wastes every contributor's time, including yours.

This is not a verdict on the work. If this change still matters:

1. rebase against current main (main has landed a lot since this PR opened),
2. confirm the diff still applies and the affected verdicts still hold,
3. reopen — the triage board picks reopened PRs up as fresh arrivals.

Wildcard holds (verdict window 2026-11-01) and Pareto land candidates are
unaffected. Thanks for the effort — high-volume repos only stay reviewable
with a queue that actually drains.