[hamzy-hams]

Fixes #1423

Problem

On main (7d63d5f), Checker::check_stmt_addr_of_loan_projections and Checker::walk_addr_of_loan_projections in compiler/crates/hy-check/src/addr_loan.rs exempt &m[k] (allow_slot_loan_root = true) in four positions where hy-emit (compiler/crates/hy-emit/src/stmt.rs) never lowers &m[k] as an O8 map-slot loan:

1. var declarations (LocalKind::Var, var p = &m["a"] / var p *acct = &m["a"]): emit_local_in_seq (stmt.rs:295-323, 371-386) only lowers LocalKind::Define | LocalKind::Let bindings with l.ty.is_none() and l.names.len() == l.values.len(), returning early for LocalKind::Var.
2. Non-identifier assignment targets (ptrs[0] = &m["a"], *pp = &m["a"]): emit_assign_in_seq (stmt.rs:1207-1213) only lowers AssignOp::Assign statements when targets.len() == values.len() and every target is PlaceKind::Ident(_).
3. Mixed-target parallel assignments (q, arr[0] = &m["a"], 1): because targets.iter().all(|t| matches!(&t.kind, PlaceKind::Ident(_))) is false when any parallel target is an index or field place, hy-emit emits &m["a"] verbatim even for the identifier target q.
4. Expr::StructLit fields on structs not rejected by A6 (struct box { p any }; b := box{p: &m["a"]}): walk_addr_of_loan_projections (addr_loan.rs:390) unconditionally skipped self.map_slot_borrow_span(st, &f.value).is_some() assuming A6 (HYB-E3103) had already fired on the struct declaration, which does not happen for any-typed or generic type-parameter fields (and continue; also skipped recursing into nested operands under Stmt::Return and Expr::StructLit).

All four shapes pass hyb check with 0 diagnostics (Verdict::Accept) and fail hyb build (go build: invalid operation: cannot take address of m["a"]), violating GUARANTEES.md Row 13 (accepted-implies-buildable) and T24 (HYB-E3104).

Fix

In compiler/crates/hy-check/src/addr_loan.rs:

• Restrict allow_root in Stmt::Local(l) (check_stmt_addr_of_loan_projections) and mark_slot_loans_local to LocalKind::Define | LocalKind::Let with l.ty.is_none() && l.names.len() == l.values.len() (preserving LocalKind::Const for HYB-E6003).
• Restrict allow_root in Stmt::Assign (check_stmt_addr_of_loan_projections) and mark_slot_loans_assign to *op == AssignOp::Assign && targets.len() == values.len() && targets.iter().all(|t| matches!(&t.kind, PlaceKind::Ident(_))), while exempting direct PlaceKind::Field root borrows that are already owned by A6 (HYB-E3103).
• In Expr::StructLit, only exempt a root &m[k] field value when HYB-E3103 has already been recorded, and recurse through walk_addr_of_loan_projections(..., allow_root) in both Stmt::Return and Expr::StructLit so nested loan projections are still checked.

Verification

• Added compiler/crates/hy-check/tests/non_lowered_slot_loan_positions.rs covering var p = &m["a"], var p *acct = &m["a"], ptrs[0] = &m["a"], q, arr[0] = &m["a"], 1, box{p: &m["a"]}, and positive controls (q := &m["a"], q = &m["a"], and A6 b.p = &m["a"]).
• cargo test --manifest-path compiler/Cargo.toml --workspace --quiet
• python3 tools/check_corpus_census.py
• python3 tools/check_conflict_markers.py
• python3 tools/check_test_registration.py
• python3 tools/corpus_harness.py --bin target/debug/hyb
• python3 tools/managed_gate.py target/debug/hyb gates/managed-baseline.json tests/checker-corpus/02-shared-map-managed.hy tests/checker-corpus/18-pool-shared-struct-managed.hy
• python3 tools/differential_loop.py --bin target/debug/hyb

[limitlessprojects]

Verified and subsumed by #1441, which we just landed on main (merge onto post-#1432 tip, battery 1775/0 · corpus tripwire clear · differential FULL 40+86) — credit-closing rather than landing both.

Your five scenarios, probed against the landed tree:

• var p = &m["k"] (var init, no type) → REJECT HYB-E3104 ✅
• var p *a = &m["k"] (typed var) → REJECT HYB-E3104 ✅
• ps[0] = &m["k"] (slice-elem store) → REJECT HYB-E3104 ✅
• q, xs[0] = &m["k"], 1 (mixed targets) → REJECT HYB-E3104 ✅
• box{p: &m["k"]} (composite field) → REJECT HYB-E3104 ✅
• loan re-loan q = &m["k"] → still ACCEPT ✅

Both PRs restrict loan marking/gating to the positions hy-emit actually lowers (define/let without type ascription, all-ident assign targets). #1441 carries the same gates plus the call-argument position coverage (52-line test file, merged verbatim), so nothing from your diff is missing. One implementation note: your struct-field early-continue keyed on self.diags.iter().any(|d| d.code == "HYB-E3103") is order-dependent on the diagnostic list — #1441's unconditional variant avoids that fragility, which also made it the safer base for the post-#1432 merge.

Credit: you and @Dimank92 (#1441) independently spotted the same non-lowered-position gap. The landed fix keeps your five repro shapes pinned green.