native_proof / typed_shape_descriptors: four lowering-independent failures (numeric-coerce on a proven read, missing typed-array fallback record, negative-divisor frem, missing layout-note marker)

[proggeramlug]

Symptom

Four perry-codegen integration-test failures on main (f05ae3b), arm64 macOS, debug, that are independent of the root-lowering default — they fail identically with and without PERRY_RS4GC=0, so they are not part of the #7370 native-roots family tracked in #7493:

tests/native_proof_buffer_views.rs

• proven_buffer_and_typed_array_reads_are_numeric_operands (:915)
  proven native numeric reads should not force JS number coercion — the emitted IR now contains call double @js_number_coerce for a read the test proves is native-numeric.
• reassigned_typed_array_store_records_runtime_fallback (:1942)
  expected reassigned typed-array store to record runtime fallback — no artifact record with expr_kind == "TypedArraySet", consumer == "TypedArraySet.slow_path", access_mode == "dynamic_fallback" and a non-null fallback_reason.

tests/native_proof_regressions/integer_modulo.rs

• integer_modulo::i32_counter_mod_unsafe_or_nonliteral_divisors_keep_frem (:75)
  negative divisor must remain on floating remainder.

tests/typed_shape_descriptors.rs

• integer_arithmetic_array_push_omits_inbounds_layout_note_and_barrier (:141)
  IR should contain block marker ….

All four reproduce when run alone, so this is not order dependence — unrelated to #7490 (poisoned mutex in typed_feedback.rs, fixed in #7492).

Why they went unnoticed

Integration suites under crates/*/tests/*.rs do not run per-PR (nightly/tag only — the CLAUDE.md pitfall), and cargo test without --no-fail-fast stops at the first failing target, so a log tail can look like a clean run.

What needs deciding

For each of the four, the question is the same and has to be answered before touching the test: did the lowering legitimately change, or did a proof regress?

• If the numeric-proof / fallback-record / layout-note behaviour changed on purpose, the assertion is stale and should be re-pointed at the current contract — and made to assert the property rather than a textual window, so the next intentional change does not silently re-break it. (fix(test): isolate typed_feedback failures — a poisoned ENV_LOCK turned one failure into five (#7490) #7492's class-field assertion is a worked example: the positional "coerce appears between these two labels" window was replaced by an end-to-end data-flow assertion.)
• If not, these are live regressions hiding behind an untested suite. An unnecessary js_number_coerce on a proven native read is a lost optimisation, but a missing dynamic_fallback record on a reassigned typed-array store is a soundness-adjacent gap in exactly the direction the test was written to catch.

Repro

cargo test -p perry-codegen --tests --no-fail-fast -- --test-threads=1