compiler-output-regression: 5 native-region-proof workloads fail their gate assertions (blocks full tier) #8261

compiler-output-regression is red in the full tier, so it blocks the release (the full tier no longer carries continue-on-error).

Repro

cargo build -p perry
python3 scripts/compiler_output_regression.py suite \
  --suite native-region-proof \
  --perry target/debug/perry \
  --benchmark-mode smoke --runs 1 --perf-counters off \
  --gate --print-summary

Current state: 6 of 11 pass, 5 fail

Failing:

• h1_native_rep_equivalence
• image_convolution
• loop_data_dependent
• numeric_arrays
• packed_f64_loop_versioning

Passing: h1_buffer_alias_negative, packed_f64_loop_versioning_negative, dynamic_fractional_array_index, loop_bound_semantics, raw_numeric_object_fields, scalar_replacement_literals.

These are real gate findings, not harness errors

Each failing workload reports "errors": [] — capture succeeds and the workload is evaluated; it fails its gate assertions about the emitted code. Per-workload detail is in the run's compiler-output-regression artifact (structural-report.json per workload, under target/compiler-output-regression/native-region-proof/<workload>/).

Why this only became visible now

Until #8097 the gate never got as far as evaluating anything: all 11 workloads died in capture with

"errors": ["PERRY_LLVM_KEEP_IR did not report a retained object path"]

because the in-process statepoint arm retained the object but never announced it, and then with FileNotFoundError: '(in-process)' because the harness executed the compile plan's placeholder clang_path as a program. Both are fixed, so the gate now measures its subject — and 5 workloads have genuine findings that were previously unreachable.

So these 5 are not new regressions from this week; they are pre-existing output regressions the gate could not report while it was broken. Someone should expect real codegen work here, not a harness repair.

Suggested triage

Start with numeric_arrays and loop_data_dependent — they are the smallest and their assertions are about numeric-region proofs, so they most directly name what changed. h1_native_rep_equivalence is the one to do last: it compares two representations, so a finding there may be a consequence of the others rather than an independent bug.