feat(#242): roster conformance for 7 models, retire the parity suite (S3, split from #367)

[Polichinel]

Epic #242 S3, split out of #367 (self-disclosed in #368). Model-side roster work only — no ensemble deletions, no ADR amendment, no register edit, no C-132 wiring. Those follow in their own PRs.

Why the parity suite is deleted here rather than separately

My earlier review proposed landing the roster first and retiring test_datafactory_parity.py afterwards, with both suites green together as the evidence. That is impossible — they assert mutually exclusive facts:

old suite	roster
:161 loss_reg == {"tobit"}	FOUNDATION sets mse
:166-170 non-empty loss_reg_sigma	test_no_tobit_sigma requires it absent
:155-157 loss_class == "focal"	FOUNDATION sets weighted_bce

Its premise is also gone: S2 (#365) migrated the viewser trio, so there are no longer two trios differing only in data source.

violet_visitor is exempt, not reconfigured

Its two config files are not touched. It is fenced by EXPERIMENT_IN_PROGRESS (e1674af9), mid-experiment, and there is uncommitted maintainer work in its config_queryset.py. The roster still declares all eight — being mid-experiment excuses a model's values, not its membership.

The fence is re-homed, not removed. test_datafactory_parity.py was the only reader of EXPERIMENT_IN_PROGRESS anywhere in the repo, so deleting it would have made the marker dead code. The set-pin comes across: EXPERIMENTS_IN_PROGRESS = {"violet_visitor"}, asserted exactly, so adding or removing an exemption is a reviewed edit — plus a non-vacuity guard, since pins over an empty parametrisation assert nothing. #367 instead asserted the marker is absent, converting an escape hatch into a prohibition.

Violet is exempt from the value pins (family, composition, gate, seed, FOUNDATION) and from the viewser-import check — it declares a datafactory source and still imports viewser, a pre-S2 leftover whose removal is an edit to a fenced config. It remains subject to every structural check: grid, targets, gate channel, region, GED features, and metadata. Those are not what a loss experiment churns.

The rusty_bucket rewiring is deferred

A consequence of the exemption, not a preference. The ensemble declares expected_samples_per_model: 16; the seven pinned members emit D×K = 4×4 = 16; violet emits 8 (D=8, no head sampler). Swapping the eight temporary_* stand-ins in now fails test_ensemble_configs.py::test_declared_modelset_and_sample_counts_match_reality — correctly. Its modelset is left as development has it, and the ensemble is rewired in the change that settles violet. The reason is recorded in the test file so the next reader does not re-derive it.

D×K

conftest.get_head_sample_count / get_produced_sample_count — K defaults to 1, so every non-family model is byte-identical — plus the four call sites that should compare against the produced width rather than D alone.

Side effect worth naming: this makes test_pfe_production_readiness's violet_visitor_calibration case pass. It was red on development because the expected width was computed as D while the artifact carries D×K.

Departures from #367

• test_roster_is_fully_pinned dropped (asserted the fence is absent)
• Fleet-wide C-132 guard and rusty_bucket's classification_targets held back — gated on views-pipeline-core#422 merging and a release being pinned here (CI installs views_pipeline_core==3.0.0 exactly, so the >=3.0.0,<4.0.0 range will not pull it)
• _load fails instead of pytest.skip for roster members — skip-on-absent is green-by-silence
• Cross-repo issue refs qualified (views-hydranet#246, views-pipeline-core#422) — bare # numbers resolve to unrelated issues here
• ensembles/rusty_bucket/configs/config_hyperparameters.py left alone — Epic #242 S3+S4: HydraNet roster reconfig + roster-conformance + D×K sample-count contract #367 collapsed it to one line, stripping 14 lines of rationale including why the pool was thinned 1024 → 128 (a 28.6 GB memory ceiling, 2026-07-20)

Verification

• ruff check . clean
• Full suite: 7747 passed, 219 skipped, 6 xfailed, 1 xpassed, 0 failed
• The single xpass (test_ensemble_guard_uses_supported_vocabulary) is strict=False, reads the installed pipeline-core rather than repo files, and is untouched by this branch
• Safety gate: no models/violet_visitor/ path in the diff