wam_rust: R−1 baseline fixes, align nested execution, cut and ITE with SWI (D129–D134)

[s243a]

Phase R−1 of the heap-cell rewrite design (docs/proposals/wam_rust_heap_cell_rewrite_design.md §12.1): fix, on today's runtime, the places where the Rust WAM target differs from SWI-Prolog, so the rewrite targets correct behaviour instead of emulating bugs. One commit per item, each with its own ledger row. A shared report is in docs/reports/wam_rust_r_minus_1_baseline_fixes.md.

item	row	fix	example: SWI vs Rust before → after
R−1f	D129	the lowered emitter's atom and fresh-variable literals now compile under the default intern feature	lowered crate didn't build (type errors) → builds and matches SWI under both feature sets
R−1a	D130	backtrack floor around the catch/3 goal and recovery, lowered call/execute/lowered_dispatch, and the dynamic rule-body solver	catch_floor: [ok] vs [] → [ok] (also lower_floor, dyn_floor)
R−1b	D131	one accumulator per active aggregate frame	nested: [[a,b]] vs [[b]] → [[a,b]]
R−1c	D132	Rust-only WAM rewrite (shared wam_target.pl unchanged): permanent variables first bound inside an ITE are initialized before the guard, limited to the code the guard reaches	ite_map(no,R): [a,b] vs [] → [a,b]; lowered and interpreter now agree on rundoite
R−1d	D133	B0 saved in the environment frame at Allocate, restored at Deallocate (standard WAM)	cut_after_call: [first] vs [first,fallback] → [first]
R−1e	D134	the catch/3 goal is opaque to cut	catch_cut was broken → [a,c]

catch/3 stays first-solution by design, so catch_cut gives [a,c] where SWI gives [a,b,c]. The harness checks it against SWI's catch(once(G),C,R) and asserts the difference explicitly.

R−1a and the hybrid wrappers: the rust_target.pl hybrid wrappers were left unchanged on purpose. Each starts with vm.reset_query(), so no older caller choice point can exist and a floor would be a no-op.

Tests

• New tests/test_wam_rust_baseline_fixes.pl (92 tests):
  • compiles each program through the Rust WAM target and compares with SWI in interpreter mode, functions mode, direct lowered calls, and a --no-default-features build;
  • every case was confirmed failing before its fix.
• Lowered exec tests and two expected-text assertions were updated for the .into() literal form.
• 6 previously failing Rust WAM plunit files now pass (cut_semantics, lowered_dispatch, lowered_ite_exec, lowered_t4/5/6). The set of failing tests is a strict subset of the base's.

Verification

• Agent, on every commit (clean export):
  • term differential 2600/0/0, corpus 51/51;
  • store differential 503/0/0, store corpus 51/51 and identical to the term corpus;
  • crate lib 260/260, CI rust conformance smoke.
• Byte identity: all resolver output JSONLs and scale --bench stdout at N=40 and N=5000 are cmp-identical to the base after every commit. The re-freeze records an unchanged baseline.
• Perf: N=40 callgrind −0.5% overall; no item above +0.05%.
• Independently re-run on the final head after a fresh build: term differential 0 divergences, corpus 51/51, crate lib 260/260, and the harness's 92/92.

Found, not fixed (documented in the report; outside R−1)

• the lowered emitter drops aggregates;
• a lowered call/execute of a runtime builtin with no label (e.g. catch/3) returns false;
• the lowered emitter emits the ','/2 goal functor as "/2";
• the \+/1 builtin arm uses a sentinel, not a floor (not reached under default options);
• catch/3 recovery is still transparent to cut;
• an aggregate frame abandoned by an exception leaves its inner list behind.

🤖 Generated with Claude Code

https://claude.ai/code/session_013PJqq7jMzces6bMwTnPbpd

---

Generated by Claude Code