[loopx-agent]

What and why

Fixes #57. FFmpeg's public result currently makes a pre-build source-policy rejection indistinguishable from a failed build or pixel-correctness failure. Record the failure stage in the producing verifier branch and read it from the structured reward when grading. The existing failed overall-score detail and summary gain a fixed, bounded explanation; raw reasons, paths, exceptions and workload details are not exposed.

Numeric scoring, validity, test counts, source checks, thresholds, workloads and rescaling are unchanged. Unknown/missing/malformed diagnostic evidence preserves the previous feedback. Positive results are unchanged. This does not fix the broad scanner's false positives, log-based score injection, root build isolation, or the old eval script's richer category hints.

Checklist for task changes

• Task version bumped to 1.0.1 for the feedback change.
• NOTES.md records the feedback-only effect and validation; unresolved known issues remain explicit.
• Same submissions, same host, interleaved old/new judges; timed scaffold run twice per image.
• No new blobs.

Old vs new judge

Old: published amd64 seededge/edgebench.judge.ffmpeg_swscale_reimplementation:e938bf6f485e, digest sha256:711b595fe8a28783b6450afd900a2d4a52fa0c409efdb78e94deed012c0ab9f6.
New: an image derived from that immutable digest, replacing exactly grade.py, run.sh, tests/compute_reward.py and tests/test.sh at their Dockerfile destinations. Baseline and toolchains are the same image layers. Each isolated container used 4 CPUs, 8 GiB, no network and the real /judge/run.sh entrypoint. This is an incremental source-image validation, not a clean rebuild of the toolchains.

Submission	Old raw score	New raw score	New diagnostic
Empty archive	0	0	Build failed / required library absent
Research JSON rejected by the existing scan	0	0	Source policy failed before compilation
Makefile with failing release target	0	0	Build failed
Library whose context constructor returns null	0	0	Correctness failed; performance not run
Original Rust scaffold, repeat 1	0.364605	0.371320	Unchanged passing result
Original Rust scaffold, repeat 2	0.371796	0.360566	Unchanged passing result

All zero-score samples stay valid with one failed test. Both scaffold repetitions stay valid with one passed test and 30/30 correctness. The speed ranges overlap; these samples do not establish statistical performance equivalence or solver improvement. No measured speedup is claimed.

The original scaffold is committed under environment/workspace/scaffold/rust; synthetic negatives need no historical solver artifacts. The source-policy sample is only a diagnostic note containing a verifier-related term; scanner rules are deliberately unchanged.

Validation: uv run --group dev pytest -q tests/ — 36 passed, 1 skipped. The new 17 cases exercise the real reward CLI → shell shim → task grader → public result parser, including absent/unknown/contradictory evidence, positive-score parity and build-output spoofing of diagnostic stage. Independent review also reran those 17 cases. Shell syntax checks pass for the two touched shell scripts.