docs: name the backend axis correctly and precise the v0.9 blocker

ROADMAP.md

@@ -39,8 +39,14 @@ no GPU, no model download, no database required.
 - [x] Backend selection via configuration (`STRUCTURA_2D_BACKEND`, `make_segmenter`).
 - [x] Geometry metric utilities (over-/under-segmentation rate, a/b-axis error;
       plus IoU/matching, precision/recall/F1) in `structura.metrics`.
-- Data-dependent: model fine-tuning and the SAM-vs-Cellpose verdict belong to
-  the evaluation milestone (v0.9), not here.
+- Data-dependent: model fine-tuning and the backend verdict belong to the
+  evaluation milestone (v0.9), not here.
+- The axis is **zero-shot generic vs. domain-finetuned**, not "SAM vs. Cellpose".
+  `CellposeSegmenter` runs Cellpose-SAM v4 (Pachitariu et al. 2025), which is
+  itself SAM-backboned — the two learned backends share an architecture and
+  differ in what they were trained on. The companion plan's H_A states it that
+  way: a specialist fine-tuned on ≤ 300 hand-labelled stones against zero-shot
+  SAM.
 
 ### v0.4.0 — 2.5D track ✅ *(released 2026-06-16)*  *(Sub-study B)*
 - [x] `dem.relief` derivatives (hillshade, slope, curvature, local relief model,
@@ -115,12 +121,23 @@ intended delivery path.
 - [ ] Validate the archaeological usefulness of the semantic ortho — Structura's
       evaluation, not Effigies'.
 
-### v0.9.0 — Evaluation harness  ⛔ *needs real trench data*
+### v0.9.0 — Evaluation harness  ⛔ *needs **annotated** trench data*
+
+The blocker is annotation, not capture. A real orthophoto and DEM exist
+(`paper/input/data/Export/`, Tiberias 2023-03-10) and the pipeline has been run
+on them end-to-end. What does not exist is ground truth: no reference stone
+outlines anywhere in the paper repo, so no metric can be scored. H_A puts a
+bound on it — ≤ 300 hand-labelled instances — which makes this a schedulable task
+rather than a wait for a season.
+
+- [ ] Ground-truth annotation on frozen hold-out scenes (**freeze the scene IDs
+      first** — plan Task 1.1; looking at model output on a scene before it is
+      designated spends it).
 - [ ] Reproducible evaluation scripts producing the paper's metrics
       (AP@IoU; completeness/correctness/quality + clDice/APLS; LoD95%).
 - [ ] Wire results into `paper/output/data-analysis/`.
 
-### v1.0.0 — Production release  ⛔ *needs real trench data*
+### v1.0.0 — Production release  ⛔ *needs the v0.9 verdict*
 - [ ] One full trench processed end-to-end.
 - [ ] DB default locked; stable public API.
 - [ ] Documentation complete; first Zenodo DOI minted.

docs/architecture.md

@@ -143,7 +143,12 @@ The active sink is chosen by `STRUCTURA_SINK` (`file` | `postgis` | `api`).
   PostGIS directly. Awaiting confirmation. The record explains why the original
   PostGIS-vs-API framing was incomplete: the API does not exist, and neither sink
   can land geometry that has no context yet.
-- **2D model choice:** SAM vs. Cellpose vs. classical CV — to be decided by the
-  comparative evaluation (blocked on real excavation data).
+- **2D model choice:** **zero-shot generic vs. domain-finetuned**, with the
+  classical watershed as the deterministic floor — to be decided by the
+  comparative evaluation. Not "SAM vs. Cellpose": `CellposeSegmenter` runs
+  Cellpose-SAM v4, which is SAM-backboned, so the two learned backends share an
+  architecture and differ in training data. Blocked on **annotated** excavation
+  data, not on capture — a real orthophoto and DEM exist and have been processed;
+  ground-truth stone outlines do not.
 *(The intake-layout question that stood here is resolved — see
 [Intake](#intake-structuraintake) above.)*