Summary

Docs-only. Writes ADR-D45 with the mechanism traced to code and data on
2026-07-10, and fixes three ledger-hygiene defects.

ADR-D45: the producer seam (rewritten)

D45 already existed (dated 2026-06-22) but framed the seam as a train/serve
value skew. On 2026-07-10 the mechanism was traced and is narrower: a
producer seam, not a value skew and not a storage problem.

• The fingerprint pins names, not values or producers.
  compute_schema_sha (feature_schema.py:253) and compute_feature_schema_sha
  (feature_schema.py:273) hash sorted column names / family membership through
  short_sha (_hashing.py:29, sha256[:12]). Neither can see a value or
  whether a producer ran.
• The six LAFA signals are declared in feature_schema.py:
  NUMERIC_FEATURES lines 113, 114, 116, 119, 120, 121; FEATURE_FAMILIES
classifier (232), self_prior (233), association (234).
• The export fills all six with a constant 0.0. _lafa_default_fields()
  (protea/core/_leaf_record_builder.py:342) returns them as 0.0; its
  docstring: the classifier_* and association_* columns "stay zero until
  later lafa-integrate slices wire their producers. A well-defined zero, not
  NaN, matching the lineage convention." (The prior note naming
  _reranker_default_fields() was wrong; that method at line 34 fills the KNN
  block, not the six LAFA columns.)
• Consequence recorded in the sealed run:
  protea-reranker-lab/results/clean_227230/comparison.json has
  feature_exclusions = "association_* and classifier_* (zero-filled in
  export) + ...".
• The live DB is fine: a 50,000-row sample of go_prediction.features on
  2026-07-10 shows 0.0% zeros for association_total and classifier_score.

Decisions adopted: no silent default (emit NaN or refuse); an export
degeneracy check (a constant declared family fails the job loudly); per-family
provenance (SignalConfig.source of the signal store). Not D45: typing the
six columns is a ~2.5 GB additive storage win (drops the ~75 GB duplicated
features JSONB) that does not close the seam. Honest scope: the
sealed champion (f_micro_w 0.4063, v227 to v230) was trained without the
classifier and association families, so the number is intact; the seam
silently removed three declared families from the model's reach.

Ledger hygiene

• index.rst intro defect: the strategic range read D1-D31 while the
  series runs to D45 (and the implementation range read 001-008 while
  009 exists). Corrected both, and clarified that D1-D31 came from master
  plan rev 3 while later records were added as the work progressed.
• Cross-repo D-number collision: PROTEA D39/D40/D41 (destructive-op
  guards, leakage-free temporal eval, lean f32 refpool) collide with
  protea-reranker-lab D39/D40/D41 (FAIR dataset packaging, IA-aligned
  training, universal booster). D34 is intentionally the same decision in
  both. Resolved by a disambiguation note rather than renumbering, because the
  PROTEA numbers (especially D40) are already cited across docs, memory, and
  thesis; qualify the lab records as lab-D39 etc.
• D38 superseded: its champion designation (prostt5 k=5 binary, Fmax
  0.7291 on v226) is superseded by the sealed board (learned encoder
  d8979601, f_micro_w 0.4063 on the v227 to v230 frame under D40 + D43).
  Marked with the existing :Superseded-by: field convention (as D09 uses) and
  Status Superseded in the index.

Verification

• check_no_em_dashes.py: OK, 0 violations.
• check_doc_refs.py: 199 broken refs before and after (all pre-existing
  autodoc :module:/:func:/:class: roles on a clean develop; this PR
  adds none).
• check_smells.py: OK, no new offenders.

Note on the task premise

The task stated D45 "has never been written down." It existed with a
mis-stated (value-skew) mechanism; this PR corrects it in place rather than
creating a duplicate. Details in the agent report.