Upload folder using huggingface_hub

harness/DATASET_SPEC.md

@@ -15,8 +15,8 @@ Status: ✅ stored · 🔧 gap being fixed · ⏳ planned stage
 | field | needed by | status |
 |---|---|---|
 | reflectance r,g,b,nir08 (160², float32, 30 m) | all | ✅ |
-| exact acquisition datetime (day, not month) | codec time encoding (1) | 🔧 fetcher saved YYYY-MM only → now YYYY-MM-DD + sidecar |
-| platform + cloud % + nodata fraction | QC, CLOUD mask tokens (1) | 🔧 sidecar JSON per scene |
+| exact acquisition datetime (day, not month) | codec time encoding (1) | ✅ YYYY-MM-DD + sidecar (v3) |
+| platform + cloud % + nodata fraction | QC, CLOUD mask tokens (1) | ✅ sidecar JSON per scene |
 | SLC-off screening | all | ✅ (platform preference + ≤5% nodata) |
 | spot tag, lat/lon, domain | stratification | ✅ manifest |
 
@@ -24,13 +24,13 @@ Status: ✅ stored · 🔧 gap being fixed · ⏳ planned stage
 | field | needed by | status |
 |---|---|---|
 | prompt point + mask (256²) | 2 | ✅ |
-| **all three SAM scales + chosen flag** | granularity edges (2) | ⏳ next harvest round |
-| per-scale QC (solidity/fragment gates at s2) | 2 | ⏳ |
+| **all three SAM scales + chosen flag** | granularity edges (2) | ✅ harvest2 (per-scale QC; s2 pass ~10%) |
+| per-scale QC (solidity/fragment gates at s2) | 2 | ✅ |
 | cloud/brightness QC flag | 1,2 | ✅ (gates from the QC audit) |
 | track id (Hungarian, cross-pass) | 1,2,4 | ✅ |
 | consensus mask per track | 1,2,4 | ✅ |
-| containment parent (laminar) | 2 | ⏳ |
-| SAM 3.1 concept + score (multiple allowed) | 3,5 | ⏳ smoke-testing; stored as INFERRED annotations, never replacing geometry |
+| containment parent (laminar) | 2 | ✅ (within-pass, chosen scale) |
+| SAM 3.1 concept + score (multiple allowed) | 3,5 | ✅ labels/labels_<tag>.npz (INFERRED; snow lat/season gate) |
 | codec entry (EFD coefs + pose + z-scored fills + clock) | 1 | ✅ derivable; stored precomputed per shard |
 
 ## Label QC protocol (before trusting SAM 3.1 at scale)
@@ -53,8 +53,7 @@ Status: ✅ stored · 🔧 gap being fixed · ⏳ planned stage
 - Task 1: sufficient once exact dates land (irregular Δt is the whole point).
 - Task 2: needs the tri-scale harvest round — single biggest gap.
 - Task 3: event bracketing is implicit in 40-year annual series for slow events;
-  fast events (floods, fires) need 2–4 extra within-event fetches per event spot
-  (monthly around known event dates) — add per-spot `event_window` to manifest.
+  fast events: DONE — 7 event_windows in manifest, 73 monthly scenes, 6 _ev shards.
 - Task 4/5: sufficient.
 
 ## SAM 3.1 labeller smoke test (2026-10-12, g6e.8xlarge, 30 m Landsat)