EPIC: the 19 target models — 11 pgm datafactory r2darts2 + 8 HydraNets — run on the latest platform

[Polichinel]

Goal (Simon, 2026-09-19): the models that matter on fimbulthul — the pgm-level r2darts2 models on the datafactory and the 8 HydraNets — run smoothly on all the latest published code: pipeline-core 3.3.0, views-r2darts2 0.2.3, views-hydranet 0.1.1.

Merge rule for this epic (Simon, 2026-09-19): a story merges into development on its own once its ritual is clean. Nothing goes to main.

The 19

family	models	today
pgm datafactory r2darts2 (11)	blue_ocean brave_heart dancing_monkey dark_necessities dark_river golden_eagle little_talks mister_bluesky old_rules red_hawk silent_fox	only on staging_202608 (#402), views-r2darts2>=0.1.0, run.sh pointing at the hydranet env
pgm datafactory HydraNet (8)	violet_visitor + 7	on development, views-hydranet~=0.1.0, no require_cuda

Definition of done — closes when every line is true on development

• 1. The 19 are on development (11 via feat(#489): the eleven pgm datafactory r2darts2 models, from staging_202608 onto development #491, 8 already).
• 2. Every r2darts2 model pins views-r2darts2[manager]>=0.2.3,<0.3.0; every HydraNet pins views-hydranet~=0.1.1 with require_cuda: True; CI pins pipeline-core 3.3.0.
• 3. Every r2darts2 and HydraNet model carries config_maturity.py (candidate); no >=0.1.0 / ~=0.1.0 pin remains.
• 4. (HydraNet half done: roster-configs-load green on 0.1.1) roster-configs-load loads all 8 HydraNets on 0.1.1; the sniffer contract accepts every one of the 19 on 3.3.0; at least one of the 11 passes calibration end-to-end on a GPU.
• 5. The catalog shows each model's data source (77 viewser / 34 datafactory / 6 synthetic).

Stories, in order

• S1 — merged 2026-09-19: chore(#478): CI on pipeline-core 3.3.0; update_readme.py survives a model whose data client is not installed #483 → 9386f753, build(deps): all 31 r2darts2 models → views-r2darts2[manager]>=0.2.3,<0.3.0 #485 → 475c901d, fix(#486): bad_romance declares regression_point_metrics — the one point forecast of 31 that did not #487 → 08cb386b (ritual: 5 agents + review-diff; 6 findings addressed; register C-81/C-89 noted)
• S2 — S2 (#488): bring the 11 pgm datafactory r2darts2 models from staging_202608 onto development #489 merged 2026-09-19 as 594b9e1b (feat(#489): the eleven pgm datafactory r2darts2 models, from staging_202608 onto development #491). Open: the end-to-end run of one of the 11 on fimbulthul (laptop OOMs at pgm) — command on S2 (#488): bring the 11 pgm datafactory r2darts2 models from staging_202608 onto development #489
• S3 — Set require_cuda: True on the 8 HydraNet configs (views-hydranet 0.1.1) #484 merged 2026-09-19 as 9a6ab6ed (feat(#484): the eight HydraNets refuse to train on CPU — require_cuda: True on views-hydranet 0.1.1 #493); C-150 registered; follow-up Fresh installs pull a torch whose CUDA build the machine's driver cannot run — the other half of views-hydranet#377 #494 (torch vs driver) outside the DoD
• S4 — S4 (#488): config_maturity.py for the 31 cm r2darts2 models — ADR-017 Phase 2, r2darts2 family #490 merged 2026-09-19 as 3f2d98ec (feat(#490): config_maturity.py for the 31 cm r2darts2 models — only the 38 stepshifter sources remain on the legacy vocabulary #496)
• S5 — Surface each model's data source (viewser | datafactory | synthetic) in the catalog — the legacy/modern split is invisible today #474 merged 2026-09-19 as 55eab03b (feat(#474): the catalog says which data source each model reaches — viewser | datafactory | synthetic #497); follow-up Render datafactory models' feature descriptions in the catalogs job — install views-datafactory there (deferred from #483 → #474 → here) #498 (datafactory descriptions in the job) outside the DoD

The ritual, per story: branch off development → build → /code-review max (5 agents) → /review-diff → address every finding → /register-risk → CI 6/6 → squash-merge → delete branch → close the story by hand → tick here.

Outside the epic (Simon's)

• The key this repo runs everything under expires 2026-11-17 12:35 (C-65) #359 — the key everything runs under expires 2026-11-17.
• Staging 202608 #402 — the rest of staging_202608 once the 11 have left it.
• Signing any of the 19 off as graduate.

Plan: ~/.claude/plans/okay-we-need-to-snuggly-wigderson.md (session views-models, 2026-09-19).

[Polichinel]

Status 2026-09-19, end of day — all five stories merged.

story	merged	ritual
S1 #483 / #485 / #487	9386f753 / 475c901d / 08cb386b	5 agents + review-diff; 6 findings addressed
S2 #489 → #491	594b9e1b	5 agents; 6 findings; C-149 registered; #492 filed
S3 #484 → #493	9a6ab6ed (+ #495 56c5c8a2 for a register-header slip I caused by merging on an UNSTABLE status)	5 angles; 4 findings; C-150; #494 filed
S4 #490 → #496	3f2d98ec	5 angles; 3 prose findings
S5 #474 → #497	55eab03b	5 angles; 5 findings; #498 filed

Definition of done: lines 1, 2, 3, 5 ✓. Line 4 is half open: roster-configs-load loads all 8 HydraNets on 0.1.1 ✓; the sniffer accepts all 19 ✓; one of the 11 pgm r2darts2 models through calibration end-to-end on a GPU — not yet. The laptop reaches datafactory fetch → audits → GPU training, then OOMs (31 GB, desktop alongside). The command for fimbulthul is on #489. When it passes, tick line 4 and close this epic.

Filed on the way, outside the DoD: #492 (prediction_frame migration for r2darts2), #494 (torch vs driver — the other half of views-hydranet#377), #498 (datafactory feature descriptions in the catalogs job). Named on #402: eight more staging models with the run.sh env bug.

Simon's, still: #359 — the key expires 2026-11-17. #402 — the rest of staging_202608. Signing any of the 19 off as graduate.