typed_f64_receiver_method_clone_raw_loads_after_composed_guards: the guard-failure edge no longer calls $generic — miscompile or intentional collapse?

[proggeramlug]

Split out of #7493, whose sweep listed this test among the failures PERRY_RS4GC=0 heals. It does not. Run alone it fails under both lowerings:

$ cargo test -p perry-codegen --test native_proof_regressions \
    typed_f64_receiver_method_clone_raw_loads_after_composed_guards -- --exact
thread '…' panicked at crates/perry-codegen/tests/native_proof_regressions.rs:11750:
receiver field or numeric arg guard failure should call the generic method body:
define double @perry_fn_typed_f64_receiver_method_ts__probe(double %arg1, double %arg2) alwaysinline {

$ PERRY_RS4GC=0 cargo test … -- --exact      # identical failure

The earlier "healed by PERRY_RS4GC=0" reading came from a whole-suite --test-threads=1 run, so it was an ordering artifact, not a lowering effect. #7493 therefore leaves this test deliberately unpinned, with a comment saying so, and hands it here.

The failing assertion

assert!(
    caller_ir.contains(&format!("call double @{generic_body}(")),
    "receiver field or numeric arg guard failure should call the generic method body:\n{caller_ir}"
);

The earlier assertions in the same test still pass — the caller does emit the method-direct guard, the raw-f64 field guard and the $typed_f64_recv call, in that order. What is missing is the $generic call on the guard-failure edge. The test's next assertion (a js_native_call_method_by_id fallback for method-direct guard failure) is not reached.

What needs deciding

Same question as #7494's four, and it has to be answered before touching the test:

• If the guard-failure edge legitimately changed — e.g. the receiver-field guard now shares the method-direct guard's fallback, so one dynamic call covers both — then the assertion is stale and should be re-pointed at the property (every guard-failure edge reaches a correct dynamic path), not at the presence of one symbol. fix(test): isolate typed_feedback failures — a poisoned ENV_LOCK turned one failure into five (#7490) #7492's class-field rewrite is the worked example: it replaced a positional window with an end-to-end data-flow assertion.
• If it did not, a guard-failure edge that no longer reaches the generic body is a miscompile: a receiver whose fields are not raw f64 would take the typed clone's raw loads anyway.

The test's own doc comment ("receiver clone must run only after method-direct and raw-f64 field guards") says the guards compose; check whether the composition still has three outcomes or has been collapsed to two.

Repro

cargo test -p perry-codegen --test native_proof_regressions \
  typed_f64_receiver_method_clone_raw_loads_after_composed_guards -- --exact --nocapture

Sibling lowering-independent failures: #7494 (four of them), #7503, #7504, #7505.