docs(register): D-01 gains independent corroboration of R9 from views-hydranet

reports/technical_risk_register.md

@@ -371,6 +371,7 @@ Maintainer decisions taken on a register concern, recorded here so the reasoning
 - **Ships with:** #64. **Shipped:** 1.1.0 published to PyPI 2026-09-18 with the change in its notes; this decision's consequence is live.
 - **Residual, now live (found by the code review of #64, 2026-09-18):** C-22's — and the direct `to_dict()` reader it warned about exists: views-pipeline-core averages group values for its WandB scalars with a mean that skips `None` but not `nan` (`views_pipeline_core/modules/wandb/utils.py`, observed on their `development` branch). One empty group makes their `AP_mean` scalar `nan`. The fix is theirs (`nanmean`), named in the release notes, and C-22's trigger is widened to cover it. **Resolved on their side 2026-09-19 (their PR #523, `e65f705`, their C-329; ships in their 3.3.0):** reproduced on 2.0.0, then fixed — nanmean semantics, `inf` still a number, an all-`nan` metric omitted. Their consumer assessment of 1.1.0/2.0.0 (tracker #73) found no defect on this side; nothing their pipeline sends is on the 1.1.0 refused-inputs list. Also, until 2.0.0: `to_dataframe()` dropped an all-`nan` column (C-40), which applied to `AP` too; the method was deprecated in 1.1.0 and removed in 2.0.0 (S7, #63; C-40 closed 2026-09-18).
 - **Versioning:** filed MINOR on the ADR-022 §1 argument that the `0.0` was never documented behaviour of this library; recorded in ADR-015 R9 so the S6 checklist can answer rule 5 against it.
+- **Independent corroboration (2026-09-30, cross-session exchange with views-hydranet).** R9 was decided on argument — three options weighed, no second implementation to appeal to. One exists: `views-hydranet`'s frozen ruler returns `nan` for average precision on an all-zero label vector (`reports/2026-07-17_lodestar_eval_dossier/tools/lodestar_score.py`, guarded at `scripts/rollout_ruler_core.py:398`), **arrived at without reference to this ruling** — that code was written in July, before R9 existed, and their session reports the behaviour was never discussed there. So an independent implementation of the same metric, on the same kind of data, exhibits the same semantics. ⚠️ **This is corroboration of the behaviour, not an endorsement of the ruling:** views-hydranet has taken no position on R9, and their session was explicit that speaking for that repo is not theirs to do. Recorded because a decision resting on argument alone is weaker than one that a second implementation independently matches, and because the converse — discovering they had chosen `0.0` — would have been evidence against.
 
 ---