Fix #201: re-fold the HRV baseline on a window switch instead of restarting calibration

Packages/StrandAnalytics/Tests/StrandAnalyticsTests/BaselinesTests.swift

@@ -205,6 +205,32 @@ final class BaselinesTests: XCTestCase {
         XCTAssertEqual(s.status, .calibrating)
     }
 
+    // MARK: - #201: HRV window switch re-folds instead of restarting calibration
+
+    /// #201: switching the HRV window re-scores the recent nights under the new window and folds the
+    /// baseline from them WITHOUT re-anchoring the epoch. An established user (≥ minNightsSeed valid
+    /// nights) therefore keeps a usable baseline centered on the new (here lower, deep-window) values
+    /// immediately — not a multi-night calibrating reset. Contrast the epoch-reset path
+    /// (testRecalibrateEpochAfterAllNightsResetsToColdStart), the old window-switch behaviour that dropped
+    /// every night to cold-start and read as "the deep-sleep setting is broken" (#195).
+    func testWindowSwitchRefoldStaysUsableWithoutEpochReset() {
+        // 21 recent nights, all re-scored under the new deep window → the lower deep-only value.
+        let days = (1...21).map { String(format: "2026-06-%02d", $0) }
+        let deepVals: [Double?] = Array(repeating: 48.0, count: 21)
+
+        // #201 path: fold with no recalibration epoch → usable, centered on the new deep value.
+        let refolded = Baselines.foldHistory(deepVals, dayKeys: days, cfg: Baselines.hrvCfg, baselineEpoch: 0)
+        XCTAssertTrue(refolded.usable)
+        XCTAssertEqual(refolded.baseline, 48.0, accuracy: 2.0)
+
+        // Old behaviour: re-anchoring the epoch to "now" (after every re-scored night) drops them all →
+        // calibrating cold-start.
+        let reset = Baselines.foldHistory(deepVals, dayKeys: days, cfg: Baselines.hrvCfg,
+                                          baselineEpoch: 4_000_000_000)
+        XCTAssertFalse(reset.usable)
+        XCTAssertEqual(reset.status, .calibrating)
+    }
+
     // MARK: - "Recalibrate Charge baseline" reset helper (noop.hrvBaselineEpoch + noop.recoveryBaselineEpoch)
 
     /// An isolated UserDefaults suite so these tests never touch the real app domain.

Strand/Screens/SettingsView.swift

@@ -906,16 +906,18 @@ struct SettingsView: View {
                     .tint(StrandPalette.accent)
                     .accessibilityLabel("HRV window")
                     .onChangeCompat(of: hrvWindowRaw) { _ in
-                        // The new window shifts every night's avgHrv, so the HRV BASELINE must re-learn or
-                        // recovery would compare the new value against a baseline still folded from the old
-                        // window (the EWMA spans further than the ~21 nights that re-score → skewed for weeks).
-                        // Re-anchor the HRV baseline to now (HRV-only sibling of "Recalibrate Charge baseline"),
-                        // then re-score + refresh so the recent trend + fresh baseline both reflect the window.
-                        UserDefaults.standard.set(Date().timeIntervalSince1970, forKey: Baselines.hrvBaselineEpochKey)
+                        // #201: the new window shifts every night's avgHrv, so the HRV baseline must reflect it
+                        // too — but a plain re-score already achieves that. analyzeRecent re-scores the recent
+                        // ~21 nights' avgHrv under the new window AND re-folds the HRV baseline from them in the
+                        // same pass, and the baseline's 14-night-half-life EWMA is dominated by that fresh
+                        // re-scored tail. So DON'T re-anchor the baseline epoch: doing so would drop all history
+                        // and force a multi-night "calibrating" reset for someone who already has plenty of nights
+                        // (that reset reading as "the setting is broken" was #195). A genuine cold-start user
+                        // (<4 valid nights) still calibrates honestly; established users see the switch immediately.
                         Task { await model.intelligence.analyzeRecent(); await model.repo.refresh() }
                     }
                 }
-                Text("Whole night is NOOP's default measure; Deep sleep pools HRV over slow-wave sleep only, reading lower and matching WHOOP. Switching re-scores recent nights and recalibrates Charge over a few nights.")
+                Text("Whole night is NOOP's default measure; Deep sleep pools HRV over slow-wave sleep only, reading lower and matching WHOOP. Switching re-scores your recent nights over the new window and takes effect right away once you have a few nights of data.")
                     .font(StrandFont.caption)
                     .foregroundStyle(StrandPalette.textTertiary)
                     .fixedSize(horizontal: false, vertical: true)

android/app/src/main/java/com/noop/ui/SettingsScreen.kt

@@ -1361,27 +1361,27 @@ fun SettingsScreen(
                         onSelect = {
                             hrvWindow = it
                             UnitPrefs.setHrvWindow(context, it)
-                            // The new window shifts every night's avgHrv, so the HRV BASELINE must re-learn or
-                            // recovery would compare the new value against a baseline still folded from the old
-                            // window (only the recent ~21 nights re-score, but the baseline EWMA spans further —
-                            // it would read skewed for weeks). Re-anchor the HRV baseline to now (same key +
-                            // mechanism as "Recalibrate Charge baseline"), then force a re-score so the recent
-                            // trend + the fresh baseline both reflect the new window. A few nights to settle.
-                            NoopPrefs.of(context).edit()
-                                .putLong(Baselines.hrvBaselineEpochKey, System.currentTimeMillis() / 1000L)
-                                .apply()
+                            // #201: the new window shifts every night's avgHrv, so the HRV baseline must reflect
+                            // it too — but a plain re-score already achieves that. analyzeRecent re-scores the
+                            // recent ~21 nights' avgHrv under the new window AND re-folds the HRV baseline from
+                            // them in the same pass, and the baseline's 14-night-half-life EWMA is dominated by
+                            // that fresh re-scored tail. So DON'T re-anchor the baseline epoch: doing so would
+                            // drop all history and force a multi-night "calibrating" reset for someone who already
+                            // has plenty of nights (that reset reading as "the setting is broken" was #195). Clear
+                            // the analyze watermark so the re-score runs even though the raw HR fingerprint is
+                            // unchanged. A genuine cold-start user (<4 valid nights) still calibrates honestly.
                             NoopPrefs.setAnalyzeWatermark(context, "")
                             vm.syncNow()
                             Toast.makeText(
                                 context,
-                                "Re-learning your HRV over the ${if (it == HrvWindow.DEEP_SLEEP) "deep-sleep" else "whole-night"} window. Charge recalibrates over the next few nights.",
+                                "Re-scoring your recent nights over the ${if (it == HrvWindow.DEEP_SLEEP) "deep-sleep" else "whole-night"} window. Charge updates as soon as it's done.",
                                 Toast.LENGTH_LONG,
                             ).show()
                         },
                     )
                 }
                 Text(
-                    "Whole night is NOOP's default measure; Deep sleep pools HRV over slow-wave sleep only, reading lower and matching WHOOP. Switching re-scores recent nights and recalibrates Charge over a few nights.",
+                    "Whole night is NOOP's default measure; Deep sleep pools HRV over slow-wave sleep only, reading lower and matching WHOOP. Switching re-scores your recent nights over the new window and takes effect right away once you have a few nights of data.",
                     style = NoopType.footnote,
                     color = Palette.textTertiary,
                 )

android/app/src/test/java/com/noop/analytics/HrvBaselineRecalibrationTest.kt

@@ -113,6 +113,34 @@ class HrvBaselineRecalibrationTest {
         assertEquals(BaselineStatus.CALIBRATING, s.status)
     }
 
+    // ── 2b. #201: HRV window switch re-folds instead of restarting calibration ───────────────────
+
+    /**
+     * #201: switching the HRV window re-scores the recent nights under the new window and folds the
+     * baseline from them WITHOUT re-anchoring the epoch. An established user (≥ minNightsSeed valid
+     * nights) keeps a usable baseline centered on the new (here lower, deep-window) values immediately —
+     * not a multi-night calibrating reset. Contrast [recalibrateEpoch_afterAllNights_resetsToColdStart],
+     * the old window-switch behaviour that dropped every night to cold-start and read as "the deep-sleep
+     * setting is broken" (#195). Parity twin of the Swift
+     * testWindowSwitchRefoldStaysUsableWithoutEpochReset.
+     */
+    @Test
+    fun windowSwitchRefold_staysUsableWithoutEpochReset() {
+        // 21 recent nights, all re-scored under the new deep window → the lower deep-only value.
+        val days = (1..21).map { "2026-06-%02d".format(it) }
+        val deepVals: List<Double?> = List(21) { 48.0 }
+
+        // #201 path: fold with no recalibration epoch → usable, centered on the new deep value.
+        val refolded = Baselines.foldHistory(deepVals, days, hrvCfg, 0.0)
+        assertTrue("established user keeps a usable baseline after a window switch", refolded.usable)
+        assertEquals(48.0, refolded.baseline, 2.0)
+
+        // Old behaviour: re-anchoring the epoch to "now" drops every re-scored night → calibrating.
+        val reset = Baselines.foldHistory(deepVals, days, hrvCfg, 4_000_000_000.0)
+        assertTrue("epoch reset drops all nights back to calibrating", !reset.usable)
+        assertEquals(BaselineStatus.CALIBRATING, reset.status)
+    }
+
     // ── 3. "Recalibrate Charge baseline" reset helper (writes BOTH epoch keys) ───────────────────
 
     /**