Build the failure views from the fused checkpoint's predictions

[duckyquang]

Summary

The taxonomy, complementarity matrix and loud-vs-silent dropout profile — the three views that are the point of src/primed_ai/failure/ — were only ever produced by the harness's own Ridge. So the project's headline contribution was demonstrated on a model none of the reported tables use, and the two disagree badly enough that it matters: the ridge harness puts drop-echo MAE at 117.5 where the fused checkpoint gives 20.55 on the same 245 rows.

This adds a --predictions route that reads the per-example dump evaluate_missing_modality.py already writes, so the views describe the checkpoint the paper reports. Nothing is refit. With two modalities the harness only ever asks for the full set and each singleton, and the eval has already scored exactly those three conditions, so it's a re-reading of the reported predictions rather than a second model:

• full → {echo, ecg}
• echo_dropped → {ecg}
• ecg_dropped → {echo}

analyze_modality_failure takes modality names directly now. It never read the embedding values, only the keys, and a caller holding predictions has no arrays to pass — list(embeddings) already did the right thing for both, so it's a rename and a docstring.

Two ways this route could quietly lie, both guarded and tested: the sanitized docs/results/ copy has its per-example blocks stripped (it raises instead of half-running), and conditions whose label vectors drift out of order would silently compare different patients. The fused report also gets its own default output dir, so it can't land on top of the ridge one.

Refs #79. I'd leave that issue open until the artifacts are regenerated — see Notes.

Notes

• Stacked on Real fairness, calibration, and baseline-gap artifacts from the canonical run #74, so the base is feat/eval-artifacts and the diff here is one commit. It depends on the provenance block and the failure_report.ridge.json naming from that PR. GitHub will retarget the base to main when Real fairness, calibration, and baseline-gap artifacts from the canonical run #74 merges.
• No artifacts were regenerated. I don't have the manifest on this machine, so docs/results/ still has no failure_report.fused.json — the export path and the naming are in place, but the file itself needs a run. That's why I'm not auto-closing [E14] Failure harness on the fused checkpoint's per-example predictions #79: its first acceptance criterion is that the three views exist for the fused checkpoint on the real test fold, and that isn't true until someone reruns.
• @kevzho since you reproduced the pipeline end to end — python scripts/run_failure_analysis.py --predictions results/missing_modality.json then python scripts/export_result_bundle.py would produce it. No existing metric should move.
• This is what [W08] Rebuild paper failure figures from real artifacts (currently unlabeled synthetic demo) #76 is waiting on. Once the fused report exists, the paper figures can be rebuilt from artifacts that match Table 4's model instead of the unlabeled synthetic demo.
• Tests 116 → 121. The new ones pin the condition→modality mapping by direction, not just by field name: the fixture makes the ECG-only arm useless, so a flipped mapping would blame echo and fail.

[sebasmos]

Read at bfd7df9 on top of b849306; 121 tests pass.

The route is right, and the case for it is stronger than the description makes it sound. The committed ridge report calls drop-echo 100 percent silent at MAE 117.5; the checkpoint calls the same condition loud at 20.55. They do not disagree on magnitude, they invert the paper's headline, so this is not a labelling tidy-up. Naming the artifact after its producer and giving the fused route its own default out dir is the correct fix, and reusing the already-scored conditions rather than refitting is what makes it a re-reading rather than a second model.

Two small things.

1. The order guard compares lvef across conditions but not ef_le_40, so a dump whose gate vector drifted would still pass.

2. Right to hold #79 open. Until failure_report.fused.json exists nothing here has touched the real fold, and #76 is waiting on that file, not on this merge.

Approving to merge after #74.

[kevzho]

The fused failure-analysis route now correctly reuses the checkpoint’s per-example predictions without refitting a model, validates condition alignment, records provenance, and includes appropriate tests. The regenerated failure_report.fused.json is committed, and the updated SHA256SUMS includes the artifact. The implementation and reproducibility requirements are complete. Approving this PR for merge.

[sebasmos]

Re-approving at 322f6a4. 121 tests pass and all eight bundle lines verify with sha256sum -c.

failure_report.fused.json is the first time the three views describe the model the paper reports, and the result is the thesis rather than a restatement of it:

Dropped	Induced critical	Silent	Silent rate	Marginal value
echo	174	39	0.224	10.12
ECG	7	6	0.857	0.81

The modality worth almost nothing on average is the one whose loss hides. That is the separation between value and detectability the framework was built to measure, now measured on the reported checkpoint rather than on planted data.

The ridge report inverts both axes, 1.0 silent for drop-echo and 0.271 for drop-ECG, so it would have told the opposite story on both counts. #76 can only be built from this file.

One thing for the paper rather than for the code: 0.857 is 6 of 7 induced-critical cases. The denominator has to travel with that rate wherever it is printed, or it reads as a stable estimate and it is not.

Worth recording that the ridge report came through @kevzho's rerun byte-identical while every torch checkpoint moved, which is what localises that nondeterminism to probe training. Provenance here is internally consistent, 7cae4f92 with the same manifest, so this artifact and the regenerated bundle describe one run.

Merging this needs #74 first, and I have asked there for one docs commit before that.