Polichinel · Sep 19, 2026 · Sep 19, 2026 · Sep 19, 2026
diff --git a/documentation/CICs/cic_evaluation_report_template.md b/documentation/CICs/cic_evaluation_report_template.md
@@ -12,7 +12,7 @@
 
 > **What is this class for?**
 
-EvaluationReportTemplate **renders** self-contained HTML evaluation reports for VIEWS forecasting models and ensembles from an **injected `EvaluationSource`** — a typed `MetricFrame` per model (ADR-018 / C-108) — rather than acquiring metrics itself at render time. It assembles the comparative metric tables, optionally renders historical-vs-predicted line graphs from on-disk predictions (parquet or numpy PredictionFrame per ADR-012), and delegates all HTML assembly to `ReportModule`. This class does **not** touch WandB: it renders purely from the injected `EvaluationSource` (a `MetricFrame` per model), which the caller (pipeline-core's reporting stage) constructs as a `MetricFrameFileSource` reading the persisted frame.
+EvaluationReportTemplate **renders** self-contained HTML evaluation reports for VIEWS forecasting models and ensembles from an **injected `EvaluationSource`** — a typed `MetricFrame` per model (ADR-018 / C-108) — rather than acquiring metrics itself at render time. It assembles the comparative metric tables, optionally renders historical-vs-predicted line graphs from on-disk predictions (parquet or numpy PredictionFrame per ADR-012), and delegates all HTML assembly to `ReportModule`. This class does **not** touch WandB: it renders purely from the injected `EvaluationSource` (a `MetricFrame` per model), which the caller (pipeline-core's reporting stage) constructs as an `EvaluationSource` (since pipeline-core 3.3.0 its `PerModelMetricFrameSource`, one `MetricFrameFileSource` per model) reading the persisted frame.
 
 ---
 
@@ -35,7 +35,7 @@ EvaluationReportTemplate **renders** self-contained HTML evaluation reports for
 - **Canonical metric selection (ADR-017).** The metric *set* shown is the **central canonical standard** in `ReportingConfig.canonical_report_metrics` (keyed by `{regression,classification}×{point,sample}`), **not** the model's own list. A model occupies a cell when its `<task>_<pred_type>_metrics` config key is present & non-empty (declared, not inferred — ADR-003); one labelled canonical table is rendered per active cell. Values come from each model's `MetricFrame` via `mean_metric_value` (the `group_id="mean"` row); a canonical metric the frame lacks (absent row or NaN) is shown with an explicit "not calculated — add `<metric>` to `<key>`" note (ADR-008), never dropped silently. **C-116:** more than one matching mean row raises `AmbiguousMetric` and renders a visible "ambiguous — multiple matching keys" cell rather than a silently-picked value.
 - **Baseline model inclusion.** Collects baseline model names from tier-specific config keys (`regression_point_baselines`, `regression_sample_baselines`, `classification_point_baselines`, `classification_sample_baselines`) and includes them in the comparative metric tables (lines 173-181).
 - **Constituent model verification.** For ensemble reports, the template asks `source.metric_frame(model)` for each declared constituent and verifies `level`/`partition` consistency across the resolved frames (via `unique_axis_value` over the frame axes), raising `ValueError` on mismatch (C-48 — a single ensemble row must not mix constituents from different levels/partitions).
-- **No run selection in the template (C-48 resolved).** The template does not select WandB runs; it only consumes `source.metric_frame(model)`. The 22/25 production blanking is resolved structurally: pipeline-core persists the evaluation-of-record `MetricFrame` and injects a `MetricFrameFileSource`, so there is no render-time run to mis-select. The interim metric-aware selection seam (`evaluation_run_resolver`) has been deleted (B2 / C-108); the same-partition stale-relog concern moves upstream to the producer (C-110, gated on #220 provenance).
+- **No run selection in the template (C-48 resolved).** The template does not select WandB runs; it only consumes `source.metric_frame(model)`. The 22/25 production blanking is resolved structurally: pipeline-core persists the evaluation-of-record `MetricFrame` and injects an `EvaluationSource` (one `MetricFrameFileSource` per model via `PerModelMetricFrameSource` since pipeline-core 3.3.0), so there is no render-time run to mis-select. The interim metric-aware selection seam (`evaluation_run_resolver`) has been deleted (B2 / C-108); the same-partition stale-relog concern moves upstream to the producer (C-110, gated on #220 provenance).
 - **Constituent resolution — degrade-and-announce (#105, honoring the #177 contract).** A declared constituent is never silently dropped. The source reporting **ABSENT** (`metric_frame(model)` returns `None`) records the model and **announces** it under "Model Metrics" as missing. A `metric_frame` call that **raises** (a **transient** failure) is **retried once**; if it still fails the model is **marked degraded** and announced distinctly (a retrieval error, not a confirmed absence). Absent and degraded models contribute no metrics row. Opt-in **strict mode** (`config["strict_constituents"] = True`) instead raises `ValueError` listing the absent/degraded models. (ADR-008 fail-loud / make-degradation-visible.) The subject model is always rendered (an absent subject shows "not calculated" cells).
 - **Metric table sorting.** Sorts combined metric DataFrames by MSLE first, then CRPS, then the first available metric (lines 265-282).
 - **Prediction sample graphs (non-fatal).** `_add_prediction_sample_graphs()` is wrapped in a `try/except` at the caller (lines 297-302). A failure in graph generation does not invalidate the metrics tables already written to the report.

diff --git a/documentation/CICs/cic_evaluation_source.md b/documentation/CICs/cic_evaluation_source.md
@@ -8,7 +8,7 @@
 
 ---
 
-> **Scope note.** Covers the `EvaluationSource` *interface* (`typing.Protocol`), its single concrete adapter (`MetricFrameFileSource` — durable, loads the persisted `MetricFrame`), the `EvaluationProvenance` value object, and the pure value queries (`mean_metric_value`, `unique_axis_value`, `AmbiguousMetric`). This is the evaluation-side counterpart of the prediction `loaders/` Ingestion surface (C-108 / #173).
+> **Scope note.** Covers the `EvaluationSource` *interface* (`typing.Protocol`), its single concrete adapter in this repo (`MetricFrameFileSource` — durable, loads the persisted `MetricFrame`), the `EvaluationProvenance` value object, and the pure value queries (`mean_metric_value`, `unique_axis_value`, `AmbiguousMetric`). This is the evaluation-side counterpart of the prediction `loaders/` Ingestion surface (C-108 / #173).
 
 Sources: `views_reporting/sources/_protocol.py`, `metric_frame_file_source.py`, `evaluation_provenance.py`, `metric_value.py`, `__init__.py`.
 
@@ -35,7 +35,7 @@ Sources: `views_reporting/sources/_protocol.py`, `metric_frame_file_source.py`,
 
 - **Port (`EvaluationSource`, a `typing.Protocol`).** Provides `metric_frame(model: str) -> MetricFrame | None` and `provenance() -> EvaluationProvenance`. A source is bound to one `target` at construction (a report is per-target), so only `model` varies per call. The render code is statically checkable against the interface (LSP/ISP); a new source is a new class, not a change here (OCP). `views_evaluation` is imported only under `TYPE_CHECKING` in the port so it stays import-light (SDP/SAP).
 - **Failure taxonomy (the #105/#177 contract).** `metric_frame` returns **`None` = absent** (no evaluation for this model → the report degrades-and-announces, never a silent drop); **raises = transient** (a retrieval hiccup → the report retries once, then marks the model degraded).
-- **`MetricFrameFileSource` (durable — the only implementation).** `metric_frame(model)` loads the persisted frame for `(model, run_type, target)` or returns `None` when its directory is absent; a corrupt/unreadable frame propagates (transient). `provenance()` is read from the subject model's frame metadata (run_id / evaluation_timestamp / data_version / scoring_code_version; no WandB url/owner).
+- **`MetricFrameFileSource` (durable — the only implementation in this repo; pipeline-core's `PerModelMetricFrameSource` composes it per model).** `metric_frame(model)` loads the persisted frame for `(model, run_type, target)` or returns `None` when its directory is absent; a corrupt/unreadable frame propagates (transient). `provenance()` is read from the subject model's frame metadata (run_id / evaluation_timestamp / data_version / scoring_code_version; no WandB url/owner).
 - **Value query (`mean_metric_value`).** Reads the `group_id="mean"` row for `(eval_type, target, metric)`: `None` when absent or NaN ("not calculated"); raises `AmbiguousMetric` on >1 matching mean row (C-116; ADR-008); else the `float`.
 - **Axis query (`unique_axis_value`).** The single distinct value of an axis (`level`/`partition`) across a frame, for the cross-constituent consistency guard; raises if the axis is non-uniform within a frame.
 - **`EvaluationProvenance`.** A frozen presentation DTO (`run_id` + optional `run_url`/`owner`/`run_date`/`data_version`/`scoring_code_version`) rendered with None-omission, so each source supplies only what it knows. `MetricFrameFileSource` populates `run_id` and the frame provenance (`data_version`, `scoring_code_version`); `run_url`/`owner` remain generic optional fields that no current source populates.
@@ -84,7 +84,10 @@ Nothing is silently dropped or silently guessed.
 ## 8. Examples of Correct Usage
 
 ```python
-# pipeline-core's reporting stage constructs the source and calls generate:
+# Direct construction (a test, a script, or a caller that owns ONE root). Since
+# pipeline-core 3.3.0 the reporting stage instead injects its
+# PerModelMetricFrameSource, which composes one of these per model, each at
+# that model's own data/generated (vpc #485 / register C-215):
 source = MetricFrameFileSource(root, run_type="calibration", target="lr_ged_sb",
                                primary_model="first_love")
 template.generate(source=source, target="lr_ged_sb")
@@ -117,7 +120,7 @@ except AmbiguousMetric:
 
 ## 11. Evolution Notes
 
-- The interim WandB scrape has been removed: pipeline-core now persists frames and injects a `MetricFrameFileSource`, the only source implementation. The render path imports no WandB, and the clickable WandB link/owner have fallen away (no source populates `run_url`/`owner`; they remain generic optional fields on `EvaluationProvenance` for a future source).
+- The interim WandB scrape has been removed: pipeline-core persists frames and injects an `EvaluationSource` built on this class — one `MetricFrameFileSource` per model since pipeline-core 3.3.0 (`PerModelMetricFrameSource`, vpc #485; register C-215 records why a single subject-rooted source blanked every comparison row before that). The render path imports no WandB, and the clickable WandB link/owner have fallen away (no source populates `run_url`/`owner`; they remain generic optional fields on `EvaluationProvenance` for a future source).
 - The on-disk layout is a locked cross-repo contract with pipeline-core's producer (C-192/C-202), executable-pinned in `tests/test_vpc_seam_contract.py`.
 - The cross-constituent consistency check moved onto frame axes (`level`/`partition`); when run-resolved partition windows are plumbed into the producer (#220), they sharpen.
 

diff --git a/pyproject.toml b/pyproject.toml
@@ -18,7 +18,12 @@ readme = "README.md"
 # (risk register C-36).
 requires-python = ">=3.11,<3.15"
 dependencies = [
-    "views-pipeline-core>=3.0.0,<4.0.0",
+    # Floor 3.3.0 (#287, 2026-09-19): the first release whose reporting stage
+    # resolves every comparison model's MetricFrame from THAT model's own
+    # data/generated (PerModelMetricFrameSource, vpc #485) — the fix for
+    # baseline/constituent rows vanishing from every ensemble report (C-215).
+    # Our seam test locks that contract; <3.3.0 stages would silently blank them.
+    "views-pipeline-core>=3.3.0,<4.0.0",
     # views-frames: the leaf data contract — PredictionFrame / TargetFrame +
     # SpatioTemporalIndex / SpatialLevel, plus the views_frames_summarize sibling
     # (MAP/HDI/quantiles), which ships in the SAME distribution (one line). numpy-only,
@@ -43,12 +48,15 @@ dependencies = [
     # `[frames]` extra survives; the only merged changes to metric_frame.py since
     # 1.0.0 are to the provenance stamp's VALUE — the `+g<sha>` suffix and S1's
     # bounding of it — which we read as opaque text). The span keeps
-    # co-resolution with views-pipeline-core across its move to ^2.0.0. A >=2.0.0
-    # FLOOR is a separate step, gated on 2.0.0 actually reaching PyPI — a floor
-    # naming an unpublished version is the C-213(a) mistake (Cluster H).
-    "views-evaluation[frames]>=1.0.0,<3.0.0",
-    # views-transformation-library removed: never imported here and already
-    # pulled transitively via views-pipeline-core (falsify P3).
+    # co-resolution with views-pipeline-core across its move to ^2.0.0.
+    # Floor 2.0.0 (2026-09-19): raised only once BOTH gates held — 2.0.0 on PyPI
+    # (2026-09-18) AND vpc 3.3.0 itself requiring >=2.0.0 (its to_dataframe
+    # egress removed) — so the shared env co-resolves. Re-qualified: suite green
+    # on 2.0.0, a 2.0.0-emitted frame round-trips through our source exactly.
+    "views-evaluation[frames]>=2.0.0,<3.0.0",
+    # views-transformation-library removed: never imported here (falsify P3);
+    # it was transitive via views-pipeline-core until 3.3.0 dropped it — now
+    # simply absent, which is correct.
     # viewser RETIRED from runtime dependencies (C-22 Resolved, epic #204): the render
     # path reads the bundled metadata parquets (metadata/data/); only
     # scripts/build_entity_metadata.py uses viewser, at dev time — see the