fix(forecast): score forecasts per pipeline, not per Core build

.changeset/forecast-bands-survive-updates.md

@@ -0,0 +1,5 @@
+---
+"ftw": patch
+---
+
+Forecast error bands and baselines now carry over across Core updates that leave forecasting alone. Before, each update started them over, so the planner's forecast margin always used its widest cold-start bands.

docs/architecture.md

@@ -255,7 +255,8 @@ state has been saved, and startup restores that saved state. The learning
 revision binds state to forecast inputs and stable hardware identities. A
 binding or input change starts fresh learning, while a compatible program
 upgrade can reuse the state. Issued forecasts use a stricter revision that also
-includes the Core build, worker bytes and pipeline policy.
+includes the worker bytes and pipeline policy, so a Core update that leaves
+forecasting alone keeps its scored errors and bands.
 
 Core keeps issued forecasts, frozen inputs, model-state references and qualified
 truth in a bounded local archive. The read-only `ftw-forecast-evaluate` source

go/cmd/ftw/forecast_site.go

@@ -46,6 +46,11 @@ type forecastIdentityReceipt struct {
 }
 
 const forecastIdentityReceiptKey = "forecast/live_identity_v1"
+
+// forecastPipelinePolicy names how Core composes the forecast it plans with.
+// The evaluation cohort keys on it rather than on the Core version, so error
+// bands and baselines survive updates that leave forecasting alone. Bump it
+// whenever Core changes what reaches the planner.
 const forecastPipelinePolicy = "energyplan-primary-v2"
 
 func newForecastSiteConfig(st *state.Store) *forecastSiteConfig {
@@ -212,7 +217,7 @@ func (s *forecastSiteConfig) RefreshIdentity(now time.Time) bool {
 		learningBase = s.accepted.LearningBaseRevision
 	}
 	learning := fmt.Sprintf("site-v2:%x", sha256.Sum256([]byte(learningBase+"/"+string(data))))
-	cohort := learning + "/" + Version + "/" + s.engineVersion + "/" + forecastPipelinePolicy
+	cohort := learning + "/" + s.engineVersion + "/" + forecastPipelinePolicy
 	revision := fmt.Sprintf("forecast-v1:%x", sha256.Sum256([]byte(cohort)))
 	opts := s.baseOptions
 	if pending && opts.HouseholdInvalidReason == "" {

go/cmd/ftw/forecast_site_test.go

@@ -437,3 +437,38 @@ func TestForecastLearningIDCarriesAcrossPortablePathPolicy(t *testing.T) {
 		t.Fatalf("receipt learning base = %q, want %q", s.accepted.LearningBaseRevision, old)
 	}
 }
+
+// Error bands and baselines are scored per evaluation cohort. A Core update
+// that leaves the site, the worker and the pipeline policy alone must keep
+// the cohort, or the bands never leave their cold-start width (#1489).
+func TestForecastEvaluationCohortSurvivesCoreUpdate(t *testing.T) {
+	st, err := state.Open(filepath.Join(t.TempDir(), "state.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer st.Close()
+	script := filepath.Join(t.TempDir(), "meter.lua")
+	if err := os.WriteFile(script, []byte("measurement code"), 0600); err != nil {
+		t.Fatal(err)
+	}
+	cfg := &config.Config{
+		Drivers: []config.Driver{{Name: "meter", Lua: script, IsSiteMeter: true}},
+		Weather: &config.Weather{Provider: "open_meteo", Latitude: 59, Longitude: 18},
+	}
+	previous := Version
+	defer func() { Version = previous }()
+	s := newForecastSiteConfig(st)
+	Version = "v0.138.0-beta.1"
+	s.Configure(cfg, nil)
+	before := s.Snapshot()
+	Version = "v0.138.1-beta.1"
+	s.Configure(cfg, nil)
+	if got := s.Snapshot().Revision; got != before.Revision {
+		t.Fatal("a Core update without forecast changes started a new evaluation cohort")
+	}
+	s.engineVersion = "another-worker-build"
+	s.Configure(cfg, nil)
+	if s.Snapshot().Revision == before.Revision {
+		t.Fatal("a new forecast worker kept the old evaluation cohort")
+	}
+}