manifests/feature_shape_canaries.json is a focused source-owned advisory suite for quickly checking whether semantic benchmark output still preserves important source shapes.
[!TIP] Run this suite before the full source-owned corpus when you want a fast first-pass answer to: “did this change break a recognizable semantic shape?”
It intentionally reuses checked-in source/binary pairs instead of adding new compiled artifacts. That keeps the canary cheap to review and safe to run on any checkout that already supports the source semantic benchmark corpus.
The suite groups existing fixtures by semantic shape rather than architecture alone:
sum_array, fill_matrix, swap)control_flow.c)llvm_smoke.c)The manifest uses explicit behavior_cases so supported side-effect functions compare observable behavior through arrays or global sinks instead of relying only on static text similarity.
python3 benchmark/source_semantic_benchmark/run_source_semantic_benchmark.py \
--manifest benchmark/source_semantic_benchmark/manifests/feature_shape_canaries.json \
--fission-bin target/release/fission_cli \
--timeout-sec 45 \
--jobs 1 \
--output-dir benchmark/artifacts/source_semantic_benchmark/feature-shape-canaries-latest
Start with:
source_semantic_summary.jsonsource_semantic_summary.mdsource_semantic_rows.jsonUseful first-pass signals:
behavior_pass_rate for observable semantic preservationstatic_semantic_score_percent for text/shape drift when dynamic behavior is unsupportedpointer-array, matrix-write, switch, loop, branch, constants, and global-sink[!IMPORTANT] This suite is advisory. It is intended to make regressions easy to see before promoting any source-semantic lane to a release-blocking gate.