diff --git a/ALGORITHMS.md b/ALGORITHMS.md
index 14db222..d487503 100644
--- a/ALGORITHMS.md
+++ b/ALGORITHMS.md
@@ -46,7 +46,6 @@ Grouped by family (subdirectory under `lib/src/onehz/`). File paths are relative
 |---|---|---|
 | `correctRr` | `foundations/rr_correction.dart` | Lipponen & Tarvainen 2019 RR artifact correction (dRR/mRR/sRR beat classification, Kubios-style) |
 | `Baselines` (Winsorized-EWMA) | `foundations/ewma_baselines.dart` | Winsorized exponentially-weighted moving baseline — the rolling personal reference most other metrics compare against |
-| PPG signal-quality index | `foundations/ppg_sqi.dart` | Skewness-based SQI |
 | inverse-variance fusion | `foundations/fusion.dart` | Standard inverse-variance weighting for combining multiple noisy estimates of the same quantity |
 
 ### `clinical/` — Tier-1 cardiac/autonomic metrics
@@ -104,7 +103,6 @@ Grouped by family (subdirectory under `lib/src/onehz/`). File paths are relative
 ### `workout/`
 | Function | File | Method | Citation |
 |---|---|---|---|
-| `detectWorkouts` | `workout/workout_detect.dart` | explicit workout detection + zones | — |
 | `autoDetectWorkouts` | `workout/auto_detect.dart` | automatic workout detection | — |
 | `hrRecovery` | `workout/hr_recovery.dart` | HRR — HR drop N seconds post-peak | Cole/Lauer 1999-style HRR |
 | `Calories.dailyEnergy` / `estimateBoutCalories` | `workout/calories.dart` | Keytel HR→kcal regression + Harris-Benedict/Mifflin BMR | Keytel et al. 2005 |
diff --git a/README.md b/README.md
index afd6d50..f9be7a6 100644
--- a/README.md
+++ b/README.md
@@ -71,7 +71,7 @@ Eight families, each its own subdirectory with its own sub-barrel, built on two
 foundation layers:
 
 - **`foundations/`** — Lipponen-Tarvainen RR artifact correction, Winsorized-EWMA rolling
-  baselines, inverse-variance fusion, a PPG signal-quality index.
+  baselines, inverse-variance fusion.
 - **`clinical/`** (Tier-1) — HRV time/frequency domain (RMSSD/SDNN/pNNx, Lomb-Scargle
   LF/HF), PRSA (deceleration/acceleration capacity), nocturnal RHR/dip, an illness-risk
   CUSUM state machine, Plews ln-RMSSD readiness, Baevsky stress index, Banister/Edwards
@@ -86,8 +86,8 @@ foundation layers:
 - **`motion/`** — ENMO/MAD activity metrics, a hybrid live/1 Hz step estimator (AN-2554
   100 Hz pedometer preferred, a gated-and-bout-length-checked 1 Hz fallback for whatever
   the live stream missed), energy-expenditure fusion.
-- **`workout/`** — workout detection (both explicit and automatic), heart-rate-reserve
-  zones, Keytel/Harris-Benedict calorie estimation.
+- **`workout/`** — automatic workout detection (bout suggestion, never explicit/
+  retroactive), heart-rate-reserve zones, Keytel/Harris-Benedict calorie estimation.
 - **`wellness/`** — the canonical composite readiness score, multivariate (Mahalanobis)
   anomaly detection, CUSUM changepoint detection, temperature-based illness flagging.
 - **`human/`** — sleep regularity index, social jetlag/chronotype, single-night event
@@ -110,7 +110,8 @@ not a feature, no matter how tempting the plausible-looking headline is.
 dart test   # run from the repo root — some fixtures resolve paths relative to it
 ```
 
-290 tests, nothing mocked — pure functions, fixture in, assertion out.
+624 passed / 6 skipped (630 total), nothing mocked — pure functions, fixture in,
+assertion out.
 
 ## If you want to add a metric
 
diff --git a/lib/src/onehz/workout/observed_max_hr.dart b/lib/src/onehz/workout/observed_max_hr.dart
index 79a08e8..7e208a6 100644
--- a/lib/src/onehz/workout/observed_max_hr.dart
+++ b/lib/src/onehz/workout/observed_max_hr.dart
@@ -143,10 +143,16 @@ Metric<HrCeiling> sessionHrCeiling(
 
   final holdMs = holdSeconds * 1000.0;
   final gapMs = maxGapSeconds * 1000.0;
+  // A real held effort's corroborating motion is not necessarily even across
+  // the hold (e.g. a couple of seconds of arm swing at each end of a quiet
+  // middle), so a start that fails the motion gate on the MINIMAL qualifying
+  // window still gets to extend further before giving up — capped, so one
+  // quiet start can't turn this into an O(n²) scan of a whole day.
+  final maxSpanMs = holdMs * 4;
   HrCeiling? best;
-  // ponytail: O(n · holdSeconds) — one session at 1 Hz, so ~15 passes over a
-  // few thousand samples. A monotonic-deque sliding minimum if it ever runs
-  // over a whole day.
+  // ponytail: O(n · holdSeconds) — one session at 1 Hz, so a bounded number of
+  // passes over a few thousand samples. A monotonic-deque sliding minimum if
+  // it ever runs over a whole day.
   for (var i = 0; i < rows.length; i++) {
     var lo = rows[i].hr;
     var motionSum = 0.0;
@@ -158,19 +164,22 @@ Metric<HrCeiling> sessionHrCeiling(
       count++;
       final span = rows[j].ts - rows[i].ts;
       if (span < holdMs) continue;
-      // The window qualifies on duration. `lo` is the bpm sustained across all
-      // of it; extending further can only lower it, so this is the best this
-      // start can do and we stop.
+      if (span > maxSpanMs) break; // gave this start its fair shot
+      // The window qualifies on duration. `lo` is the bpm sustained across
+      // all of it. Only stop once the motion actually corroborates — that's
+      // the earliest point extending further can only lower `lo` for no gain.
       final motion = motionSum / count;
-      if (motion >= gate && (best == null || lo > best.bpm)) {
-        best = HrCeiling(
-          bpm: lo,
-          tsMs: rows[i].ts,
-          heldSeconds: (span / 1000).round(),
-          motionG: motion,
-        );
+      if (motion >= gate) {
+        if (best == null || lo > best.bpm) {
+          best = HrCeiling(
+            bpm: lo,
+            tsMs: rows[i].ts,
+            heldSeconds: (span / 1000).round(),
+            motionG: motion,
+          );
+        }
+        break;
       }
-      break;
     }
   }