chore(naming): cross-repo regression-target consistency — companion coordination (views-models portion superseded by #154)

[Polichinel]

Status (re-scoped 2026-06-26)

The views-models portion of this issue is done/superseded — this issue now tracks only the cross-repo companions.

• The 11-straggler rename shipped in chore(naming): standardize regression target to lr_ged_sb in 11 models (#151) #153 (merged): those models were made consistent to lr_ged_{sb,ns,os}.
• EPIC EPIC: Make views-models agnostic to the regression-target name #154 then reframed the approach: instead of forcing every model to one name, views-models is now agnostic to the target name — config is the single source of truth (get_regression_targets() accessor, ADR-013). views-models no longer depends on any model sharing a name, and its tests/tooling carry zero load-bearing target-name literals.

So the original "standardize across all models" goal is replaced by derivation. What remains here is cross-repo companion coordination for global naming consistency. These are maintainer-owned and do not block views-models (its suite is name-agnostic).

Companion coordination (cross-repo)

• views-pipeline-core — scaffold template seeds ln_sb_best/lr_sb_best (the root cause of drift); fix the seed so new models don't regenerate the inconsistency. The one essential companion. Related naming-coupling deps from EPIC: Make views-models agnostic to the regression-target name #154: pipeline-core docs(reconcile): repoint CountryMapping docstring to views_frames_reconcile #203 (heterogeneous pooling), Environment lifecycle: run.sh auto-created conda envs silently run stale deps (reports 'up-to-date' while running old code) #204 (decouple artifact filenames from the target name), Hotfix: add target_transform to pink_ponyclub stepshifter models (run against released stack) #205 (DL training reads merged config).
• views-faoapi — plotting default already lr_ged_sb ✓; fixtures only.
• views-datafactory — consumer-bridge rename map emits lr_*_best.
• views-postprocessing — fixtures (Layer-2 pred_ln_* delivery schema excluded).
• views-hydranet — ~327 fixture/config occurrences of lr_sb_best.
• views-reporting — demo script + fixtures.

Other repos' tests are self-contained — they do not break from views-models. The companions are for global consistency, except the pipeline-core template seed, which is the one fix that prevents the inconsistency regenerating.

Two naming layers (context — don't conflate)

• Layer 1 — model target / actuals: lr_ged_sb canonical. The framework treats the target name as an arbitrary string (no logic branches on it, verified across repos), so renaming is safe.
• Layer 2 — forecast delivery columns: pred_ln_sb_best / pred_ln_sb_prob, hardcoded in views-postprocessing/.../unfao/managers/unfao.py. Separate delivery schema — out of scope.

Out of scope

Layer-2 delivery schema (pred_ln_*); upstream UCDP/viewser/datafactory source columns (ged_sb_best* — not ours to rename).

History

Original goal: force every model to lr_ged_{sb,ns,os}. #153 renamed the 11 stragglers; #152 (ranger rename) closed as moot. EPIC #154 reframed to agnosticism (ADR-013), closing the views-models scope and leaving this issue as the companion tracker.

[Polichinel]

Companion issues created (2026-06-24)

• views-pipeline-core#201 — scaffold template (essential) + fixtures
• views-faoapi#99 — plotting default (already ✓) + fixtures
• views-datafactory#249 — consumer-bridge lr_*_best reconciliation
• views-postprocessing#49 — fixtures (Layer-2 pred_ln_* excluded)
• views-hydranet#136 — 327 lr_sb_best fixture/config occurrences
• views-reporting#166 — demo script + fixtures

Division of labour: views-models (#151) handled here; the 6 companions handled in their own repos.

[Polichinel]

views-models half in progress — PR #153 (base development) standardizes the 11 config-only-safe stepshifter stragglers to lr_ged_sb and adds a beige guard test enforcing canonical targets going forward.

Scope refinement from investigation: the guard surfaced that the true non-canonical set is larger than 11. Breakdown:

• 11 stragglers — done in chore(naming): standardize regression target to lr_ged_sb in 11 models (#151) #153 (config-only, no artifacts, no retrain).
• 6 synthetic *_dream models (synth_target) — intentionally synthetic, exempt.
• 4 ranger baselines (black/green/pink/yellow, lr_ns_best/lr_os_best) — split to chore(naming): standardize ranger ns/os targets to lr_ged_* (needs prediction regeneration) #152: they have stored predictions keyed by the old name, so they need a rename + regeneration, not a config-only flip.

Companions in the 6 other repos (#201/#99/#249/#49/#136/#166) remain as tracked.

[Polichinel]

✅ views-models half merged to development via #153 (merge commit 0b6a1cf): the 11 config-only-safe stragglers now use lr_ged_sb, plus a beige guard enforcing canonical regression_targets. Rides to main with the next development→main promotion.

Remaining under #151: the 6 companion repos (#201/#99/#249/#49/#136/#166) and the ranger ns/os rename + regeneration (#152).

[Polichinel]

Re-scope note. The views-models portion of this standardization is superseded by EPIC #154 (regression-target name agnosticism — config is the single source of truth, nothing hardcodes the name). The merged #153 (11 models on lr_ged_sb) stays as harmless consistency, but no further model renames are pursued; instead views-models becomes name-agnostic (tracking #162). The 6 companion repo issues (#201/#99/#249/#49/#136/#166) remain maintainer-owned and should be re-framed from 'rename' to 'de-hardcode / be agnostic' where applicable.