[wilke]

StabiliNNator ships in the same container as PredictStructure but had zero matrix coverage, so 59/59 passed silently meant one of the two apps. Its first ever real submission failed at dispatch — tasks 23450684 and 23450690 — long after folding_260821.3 was deployed and reported green.

• submit_job takes a per-case app key, defaulting to PredictStructure
• S01/S02/S03 cover analysis_type both / proline / disulfide
• the --include-alphafold filter no longer assumes every case has a tool key (it crashed with KeyError: 'tool' on the first non-PredictStructure case)
• 1crn_small.pdb staged into the workspace inputs/ folder so the cases follow the existing path convention

S01 currently FAILS against production, and that is the correct result — it reports the partition => 'normal' dispatch bug that the matrix previously could not see. It should stay red until App-StabiliNNator.pl is fixed.

Companion change in runtime_build (56b46ff): the acceptance harness now asserts the content of policy_data.partition rather than just rc=0, and goes red on the current container while PredictStructure passes the same check.

612 passed, 10 skipped.

[wilke]

Updated — the cases now assert delivered outputs, not just status == completed.

Running S01-S03 against folding_260824.2 exposed why that matters. All three jobs completed on coconut in 12-16 s with success: 1, and the original assertion reported PASS. But the job folders were empty and output_files was []: StabiliNNator uploads its results flat into output_path instead of the job folder (filed as CEPI-dxkb/stabiliNNatorApp#18). The report exists — a 611 KB 1crn_small_report.html — just not where the workspace or the UI can find it.

That is the same weakness I flagged in the acceptance harness earlier: asserting the exit code rather than the effect.

• judge() fails a completed job whose output_files is empty, and honours an optional expect_output suffix — S01-S03 require *_report.html
• fetch_output_files reads the job object with p3-cat, falling back to the folding container because p3-cat is not on the host PATH. Without that fallback the helper returned None on every call and the assertion silently never ran — which is exactly what happened on my first attempt, and S01 passed again for the wrong reason
• unreadable (None) is kept distinct from empty ([]), so a workspace hiccup is not reported as a missing-output bug

Verified both directions: S01 now FAILs with completed but no output_files; E01 and B01 still PASS with real outputs. 612 unit tests pass.

Still holding this PR: S01-S03 are red, correctly, until stabiliNNatorApp#18 is fixed.