Clarify scientific projection telemetry scope

Author: charlesmartin14

docs/WWPGD_PROJECTION_DOSE.md

@@ -1,6 +1,12 @@
 # WW-PGD projection dose telemetry
 
-Projection rows report the realized intervention size and distinguish scheduled events from actual per-layer applications.
+The main scientific runner’s `wwpgd_projection.csv` rows report the realized intervention size and distinguish scheduled events from actual per-layer applications.
+
+## Scope
+
+This schema is applied by the shared durable artifact writer used by `run_scientific_single`. It covers the scientific Level 0–2 WW-PGD, delayed-onset, adaptive, and norm-matched-sham arms that write the root experiment artifact `wwpgd_projection.csv`.
+
+It does not silently redefine the separate invalid-for-science `run_single` smoke artifact or the isolated `level_0_wwpgd` package’s independent projection schema. Those paths have different lifecycle contracts and require separate schema changes if they are ever promoted to scientific evidence.
 
 ## Realized dose
 
@@ -21,7 +27,7 @@ This is an observed dose. It is not `blend_eta`, requested hardness, or a config
 
 ## Scheduled event indexing
 
-`projection_event` is zero-based. Rows also include:
+Within the main scientific runner, `projection_event` is zero-based. Rows also include:
 
 - `projection_event_index_base = 0`
 - `projection_event_number = projection_event + 1`
@@ -36,8 +42,8 @@ Actual applications are counted independently for each `layer_name` from the dur
 - `layer_application_index`: one-based application count for an applied row; empty when no dose was applied
 - `is_first_applied_projection`: true exactly when `layer_application_index == 1`
 
-The count is reconstructed from the committed projection-row history before every append. It therefore remains correct after checkpoint resume or process restart without adding process-local counters to the optimizer, controller, model, or checkpoint state.
+The count is reconstructed from the durable projection-row history before every append. Checkpoint recovery first reconciles the CSV to its committed prefix, so the count remains correct after resume or process restart without adding process-local counters to the optimizer, controller, model, or checkpoint state.
 
 ## Compatibility
 
-These fields extend the projection CSV schema but do not change model weights, optimizer state, projection mathematics, spectral targets, or defaults. An incomplete run created with an older projection CSV header is rejected before append; start a fresh run under the new schema instead.
+These fields extend the main scientific projection CSV schema but do not change model weights, optimizer state, projection mathematics, spectral targets, or defaults. An incomplete scientific run created with an older projection CSV header is rejected before append; start a fresh run under the new schema instead.