Epic #242 S3+S4: HydraNet roster reconfig + roster-conformance + D×K sample-count contract

[Polichinel]

Epic #242 story S3 (#246), with the S4 sample-count contract reconciliation folded in (it's the only path to a green suite — the roster's D×K sampler trips the old contract).

S3 — roster + tidy-up

• Reconfigure the 8 HydraNets to the LOCKED roster (pre-registration 05, 2026-08-08):
  gated_NB (nb, soft_gate: violet 42 / bright_starship 43 / bold_comet 44),
  th_gated_NB (nb, threshold_gate 0.5: blazing_meteor 45 / heavy_freighter 46),
  mixture_NB (mixture_nb, soft_gate: pink_pirate 42 / blue_stranger 43 / purple_alien 44).
  All share the v2 foundation (mse · softplus · body_supervision=all · weighted_bce pw2 ·
  rollout_feedback=sample · bn_recalibrate · D×K=4×4 · lr_/by_ targets), all datafactory /
  africa_me_legacy. rusty_bucket points at all 8 (concat, 8×16=128) and declares
  classification_targets for the C-132 gate pooling.
• Retire golden_hour + stellar_horizon — the two 3-member C-47 parity ensembles, superseded
  by the 8-member rusty_bucket (the viewser↔datafactory comparison they existed for is moot; S2
  migrated all three viewser models to datafactory).
• Replace test_datafactory_parity.py (trio-mirror, tobit) with test_roster_conformance.py:
  pins each model to its roster family + the shared v2 foundation + grid/targets + datafactory,
  and the rusty_bucket ensemble incl. the C-132 gate channel + no stale targets key.
  EXPERIMENT_IN_PROGRESS is now unused (violet is pinned, not skipped).
• Downstream consequence fixes: stale-loss docstring guard repointed; delivery-map shadow
  count 117→115; env-sharing pin views_ensemble 13→11.

S4 — sample-count contract (ADR-015 §6)

An ADR-067 family head emits D×K = n_posterior_samples × n_head_samples = 16 draws per cell,
not D=4. The contract equated "samples per model" with n_posterior_samples alone, so every
family model's on-disk width (16) and rusty_bucket's expected_samples_per_model (16) tripped it
(9 red tests). Fix: a shared conftest.get_produced_sample_count (D×K; K defaults to 1 for
non-family models — pre-#242 behaviour byte-identical), routed through the ensemble contract, the
PF output-width checks, and the drift report. ADR-015 §6 records the reconciliation.

Register

C-132 updated with the durable fix (see cross-repo PR below); C-71 noted the roster landed and
flags Cluster E (C-47/C-71/C-72/C-74/C-87) for a /review-rr pass.

Verification

Full suite: 7574 passed, 0 failed; ruff clean. No prediction artifacts or the transient
gate-hack committed.

⚠️ Merge-together dependency

The classification_targets declaration only actually pools the gate once the framework fix lands:
views-platform/views-pipeline-core#422 (C-132). Merge these two together.

🤖 Generated with Claude Code

[Polichinel]

⚠️ Disclosure: this PR was authored by an AI agent. Please read the self-critical review notes (cross-repo dependency, un-ratified upstream ADR-067, governance edits) before evaluating it: #368.

[Polichinel]

Superseded in part. Two pieces of this are now on development, split per the review in #368's thread:

• feat(#242): roster conformance for 7 models, retire the parity suite (S3, split from #367) #369 — the roster configs for 7 models, the D×K plumbing in conftest, and the retirement of test_datafactory_parity.py. Merged 43a9f152.
• docs(ADR-015): §6 — the count that matters is the count a model produces (S4, split from #367) #370 — ADR-015 §6 and the C-91 restatement. Merged 2b13bf02.

Three deliberate departures, each with a reason:

1. test_roster_is_fully_pinned was dropped. It asserts EXPERIMENT_IN_PROGRESS is absent, which converts an escape hatch into a prohibition. test_datafactory_parity.py was the only reader of that marker anywhere in the repo, so deleting the file would have removed the fence silently. feat(#242): roster conformance for 7 models, retire the parity suite (S3, split from #367) #369 re-homes it as a set-pin, so adding or removing an exemption is a reviewed edit.
2. violet_visitor was not reconfigured. It is fenced, mid-experiment, and there is uncommitted maintainer work in its config_queryset.py. It is exempt from the value pins and the source-migration checks, and remains subject to every structural check.
3. The rusty_bucket rewiring is deferred — a consequence of (2), not a preference. The ensemble declares expected_samples_per_model: 16; the seven pinned members emit D×K = 16; violet emits 8. Rewiring now fails the ADR-015 contract test, correctly.

Two findings from reviewing this that are not in #368:

• Deleting golden_hour/stellar_horizon leaves investigations/compare_parity.py:26 — ENSEMBLE_PAIR = ("golden_hour", "stellar_horizon") — pointing at removed directories. investigations/ is not under CI, so the five green checks say nothing about it.
• ensembles/golden_hour/reports/preanalysis_plan_calibration.md (208 lines) is the sole evidential basis for register entries C-44, C-45, C-46, C-47. This PR deletes it.

Still to come from this branch: the two ensembles are being retired to README-only reusable slots rather than deleted outright (maintainer's call — they become slots for future HydraNet ensembles), with the preanalysis plan archived first. The C-132 work waits on views-pipeline-core#422 merging and a release being pinned here.

Closing this in favour of the split is reasonable once the remaining pieces land; leaving it open as the reference is also fine. Your call.

[Polichinel]

Finding from the views-pipeline-core side, on the deferred rusty_bucket scope. Not a request — your call.

The classification_targets addition to ensembles/rusty_bucket/configs/config_meta.py will be rejected by CoreConfigSniffer at config load, and this is independent of whether pipeline-core#422 merges.

Run against the real sniffer, three variants:

today (regression only)              : PASSES
this PR's config as written          : RAISES
  -> "classification_targets is non-empty but none of
      {classification_point_metrics, classification_sample_metrics}
      are present. Add at least one classification metric key."
this PR's config + a classification metric : PASSES

_check_targets_and_metrics runs inside sniff_all, and both ensemble managers call it before any side effect (prediction_frame_ensemble.py:328, ensemble.py:158). So this is a load-time refusal, not a runtime surprise.

Suggested fix, entirely your call: add a classification metric key alongside the targets — e.g. "classification_sample_metrics": ["AP"] — since AP is the metric the gate-pooling work is measured on anyway. Whether AP is the right choice is your judgement; the sniffer only requires that some classification metric is declared when classification targets are.

Why I am flagging it now rather than after #422 merges: the discussion on this PR and on pipeline-core#427 both characterise the pipeline-core change as the blocker and this config as inert until it lands. That is half right. The pipeline-core change is needed for the gate to actually be pooled — but the config as written does not get that far, because it does not load. Merging #422 first would not reveal it; the config would still be refused.

One correction to something I said on pipeline-core#427, for the record: I claimed the failure mode here would be a silent wrong answer — the gate quietly dropped. It is not silent. The sniffer refuses loudly, which is the better outcome and means this is a smaller problem than I described.

Separately: the gate-pooling effect itself is real and measured — views-hydranet's EXP-03 log records h1 AP moving 0.316 → 0.456 (sb), 0.177 → 0.355 (ns), 0.135 → 0.225 (os) once the gate is pooled. I had earlier suggested #422 could not deliver the improvement it claimed; that was wrong, and the measurement is what changed my mind.

---

Verified 2026-08-10 by running CoreConfigSniffer._check_targets_and_metrics against all three config shapes. Cross-ref: pipeline-core #422, #427, core_config_sniffer.py:261; your #368; views-models#371 (the CI gap that would have caught this).