plan(signal-store): D45 is a producer seam, not a storage seam

[frapercan]

Track B was about to be implemented on a false premise. This corrects the plan before any migration is written.

What the plan said

The six scalars (classifier_score, self_prior_score, association_cross, association_total, classifier_present, association_present) live in the go_prediction.features jsonb outside feature_schema_sha, so their values can drift unnoticed; promoting them to typed columns closes D45.

What the code and the data say

1. The six are already inside the fingerprint. feature_schema.py declares all of them in NUMERIC_FEATURES and in the classifier, self_prior and association families. compute_feature_schema_sha(families, drop) hashes "<family>=<sorted columns>", i.e. the declared schema. It covers the names.
2. What it cannot cover is whether a producer ran. A family whose columns are all present and all zero produces exactly the same sha.
3. The export fills three declared families with a constant 0.0. protea/core/_leaf_record_builder.py::_reranker_default_fields(), whose own docstring reads: "the classifier_* and association_* columns stay zero until later lafa-integrate slices wire their producers. A well-defined zero, not NaN." Only self_prior_score is overwritten, under the compute_self_prior flag.
4. The sealed run recorded the consequence. results/clean_227230/comparison.json: feature_exclusions: "association_* and classifier_* (zero-filled in export)".
5. The database is fine. A 50,000-row sample of go_prediction.features finds 0.0% zeros for association_total and 0.0% for classifier_score. The predict path writes real values; the export builds records from scratch and invents zeros.

Consequence

D45 is a producer seam. Three declared families shipped semantically null and every fingerprint matched. Closing it means a failing export, not a column type:

• no silent default (emit NaN, which LightGBM reads as missing, or refuse the row);
• a degeneracy check at export (a declared family constant across a shard fails the job);
• provenance per family (which producer, at which version, over which snapshot). That is SignalConfig.source, and it is why the store is the backend of explainability rather than a disk optimisation.

The six typed columns (~2.5 GB) and the resulting features blob drop (~75 GB of go_prediction's 101 GB; 54 of its 60 keys duplicate typed columns) remain worth doing, verified against the sealed numbers first, as a separate storage-motivated step under the additive-only invariant.

Not verified, and not claimed

No local parquet carries these columns (the bench and reranker-vnext datasets predate the schema). The sealed dataset lives in MinIO and was not read. The three lines of evidence above stand on their own; the parquet would be a fourth.

Files

• SIGNAL-STORE.md: new section 1.2 (the seam, traced); section 1 and 1.1 corrected; step 0 rewritten.
• AUTONOMOUS-RUN.md: Track B preamble and step 0 corrected.