Fix all 12 audit issues from fan-first calibration review

Major — documentation rewritten for fan-first approach:
- README calibration section: describes stabilization sweep, not binary search
- README modes table: shows stabilization window and max wait, not iterations
- ROADMAP calibration section: describes fan-first with control curve formula
- ROADMAP header: calibration redesign marked complete

Moderate:
- CalibrationMode.description timing corrected (17/25/35 min)
- CLI print: "5 fan speed levels" (was 4)

Code quality:
- Removed dead performanceCeiling constant (84°C ceilingTemp is used)
- Fixed misplaced doc comment on isStabilized
- Removed orphaned doc comment
- Fixed Metal resource leak in CalibrationView (nil out on stop)
- CalibrationMode doc comments: match actual maxWaitPerLevel values

Author: Producer Guy

README.md

@@ -138,47 +138,45 @@ Apple doesn't do this because silence sells in store demos and most users never
 For best results, calibrate Smart to your specific machine:
 
 ```bash
-sudo thermalforge calibrate                    # Standard (~25 min)
-sudo thermalforge calibrate --mode quick       # Quick (~15 min)
-sudo thermalforge calibrate --mode optimized   # Optimized (~40 min)
+sudo thermalforge calibrate                    # Standard (up to 25 min)
+sudo thermalforge calibrate --mode quick       # Quick (up to 17 min)
+sudo thermalforge calibrate --mode optimized   # Optimized (up to 35 min)
 ```
 
-Calibration answers one question per temperature point: **what fan speed holds this machine at this temperature?**
+Calibration answers one question per fan speed: **what temperature does this machine stabilize at with fans at X%?**
 
 It works in three phases:
 
 1. **Discover intensity** — finds the stress level that heats your machine at ~1°C/sec (matching real workloads, not synthetic maximum). Uses Metal compute + CPU stress.
-2. **Heat to target** — runs stress with fans on auto, lets the machine warm naturally to each target temperature (60°C, 65°C, 70°C, 75°C, 80°C, 85°C).
-3. **Find holding speed** — at each target, binary-searches for the fan speed that stabilizes temperature there. Tests a fan speed, measures whether temp is rising/falling/stable, adjusts up or down until it converges.
+2. **Fan-level stabilization sweep** — sets fans to 100%, applies calibrated stress, and waits for temperature to stabilize. Then reduces fans to 80% and waits again. Repeats at 60%, 45%, and minimum. At each level, the equilibrium temperature is recorded.
+3. **Build control curve** — transforms the raw data (higher fan = lower equilibrium temp) into a control curve (higher temp = more fan needed) that Smart reads at runtime.
 
 The result is a temperature-to-fan-speed lookup table specific to your machine:
 
 ```
-60°C → 30% fans
-65°C → 42% fans
-70°C → 55% fans
-75°C → 68% fans
-80°C → 82% fans
-85°C → 96% fans
+60°C → 39% fans
+65°C → 53% fans
+70°C → 65% fans
+75°C → 78% fans
+80°C → 90% fans
+85°C → 100% fans
 ```
 
-Smart reads this table and interpolates for any temperature. At 72°C on your machine, Smart knows it needs ~58% fan speed — not a guess, a measurement.
+Smart reads this table and interpolates for any temperature. At 72°C on your machine, Smart knows it needs ~70% fan speed — not a guess, a measurement.
 
-The machine never exceeds the target temperature because calibration controls the target, not the fan speed. A 95°C safety backstop is always active as a failsafe.
+Temperature is protected by three layers: an 84°C ceiling (skips lower fan levels), a 90°C safety stop (maxes fans immediately), and a 95°C backstop (always active regardless of what's running).
 
 ### Calibration modes
 
-Modes control how many binary search iterations per target and how long each test holds:
+Modes control how long the machine waits for thermal equilibrium at each fan level:
 
-| Mode | Time | Binary search iterations | Hold time per test |
+| Mode | Time | Stabilization window | Max wait per level |
 |---|---|---|---|
-| **Quick** | ~15 min | 4 per target | 8 seconds |
-| **Standard** | ~25 min | 6 per target | 15 seconds |
-| **Optimized** | ~40 min | 8 per target | 25 seconds |
+| **Quick** | up to 17 min | 60 seconds (30 readings) | 2.5 minutes |
+| **Standard** | up to 25 min | 90 seconds (45 readings) | 4 minutes |
+| **Optimized** | up to 35 min | 120 seconds (60 readings) | 6 minutes |
 
-More iterations and longer hold times produce more accurate data. Standard is recommended for most users.
-
-Timing is based on measured thermal time constants of 90-120 seconds for Apple Silicon laptop heatsink assemblies (Notebookcheck M1-M4 MacBook Pro stress tests, [Max Tech](https://www.youtube.com/@MaxTech) sustained performance testing). Mac Studio's larger thermal mass (~2-3x) is covered by Standard mode's timing.
+Longer stabilization windows produce more accurate equilibrium measurements. Standard is recommended for most users. Timing is based on measured thermal time constants of 90-120 seconds for Apple Silicon laptop heatsink assemblies (Notebookcheck M1-M4 MacBook Pro stress tests, [Max Tech](https://www.youtube.com/@MaxTech) sustained performance testing).
 
 **Smart works without calibration** — it uses a conservative default curve. Calibration makes it precise for your hardware.
 

ROADMAP.md

@@ -15,15 +15,15 @@ Proactive thermal curve that monitors temperature velocity and ramps fans before
 
 ### Calibration (Built)
 
-Temp-first machine-specific thermal profiling. Heats to target temperatures, binary-searches for the fan speed that holds each one.
-
-- Adaptive intensity finder: discovers the ~1°C/sec stress level for this machine before calibration
-- 6 target temperatures: 60, 65, 70, 75, 80, 85°C
-- At each target: binary search for holding fan speed (4-8 iterations per target depending on mode)
-- Produces temp→fan lookup table: "60°C needs 30% fans, 70°C needs 55% fans, ..."
-- Machine never exceeds target temp — we control the target, not the fan speed
-- 95°C safety backstop always active during both heating and search phases
-- Three modes: Quick (~15 min), Standard (~25 min), Optimized (~40 min)
+Fan-first stabilization sweep. Sets fan speed, applies calibrated stress, waits for thermal equilibrium, records the stabilization temperature.
+
+- Adaptive intensity finder: discovers the ~1°C/sec stress level for this machine
+- 5 fan levels swept high to low: 100% → 80% → 60% → 45% → minimum
+- At each level: wait for temperature to stabilize (stdev < 0.5°C AND slope < 0.05°C/sec)
+- Raw data transformed to monotonically increasing control curve via `fan_control(T) = (1.0 + minPct) - F_equil(T)`
+- Smart reads the curve via `fanPercentForTemp()` with interpolation
+- Protection: 84°C ceiling (skip lower levels), 90°C safety (max fans), 95°C backstop (always active)
+- Three modes: Quick (up to 17 min), Standard (up to 25 min), Optimized (up to 35 min)
 - CPU+GPU combined, CPU only, or GPU only stress types
 - Downgrade prevention: Quick can't overwrite Standard or Optimized
 - In-app UI: mode picker, progress bar, live temp, stop button
@@ -61,7 +61,7 @@ Research-grade data export: `thermalforge log`
 
 ---
 
-## Profile + Smart Redesign (Complete) + Calibration Redesign (Pending)
+## Profile + Smart + Calibration Redesign (Complete)
 
 The entire profile system, Smart curve, and calibration need to be redesigned as one cohesive system. Current profiles are binary switches that immediately set fans to a fixed percentage. They should be proportional curves that respect Apple's fan hardware behavior.
 

Sources/ThermalForgeApp/CalibrationView.swift

@@ -64,6 +64,12 @@ final class CalibrationState: ObservableObject {
         activeStressFlag?.stop()
         activeStressFlag = nil
 
+        // Release Metal resources
+        gpuPipeline = nil
+        gpuQueue = nil
+        gpuBuffer = nil
+        gpuElementCount = 0
+
         // Cancel async tasks
         task?.cancel()
         timerTask?.cancel()

Sources/ThermalForgeCore/Calibration.swift

@@ -107,8 +107,6 @@ public struct CalibrationData: Codable {
     }
 
     public var isValid: Bool { validationError == nil }
-
-    /// Interpolate the RPM percentage needed to hold a target temperature under load.
 }
 
 // MARK: - Persistence
@@ -167,23 +165,23 @@ extension CalibrationData {
 ///   ~20-50 J/K × 0.3-0.8 K/W = 60-180s for laptop heatsink assemblies
 /// - 3 time constants = 95% of steady state, 5 time constants = 99.3%
 public enum CalibrationMode: String, CaseIterable {
-    /// 5 fan levels × ~100s each + intensity finding + cooldowns ≈ 15 min
+    /// 5 fan levels × up to 2.5 min each + intensity finding + cooldowns
     /// 60-second stabilization window (~80% accuracy)
     case quick
 
-    /// 5 fan levels × ~150s each + overhead ≈ 22 min
+    /// 5 fan levels × up to 4 min each + overhead
     /// 90-second window (near one time constant, ~90% accuracy)
     case standard
 
-    /// 5 fan levels × ~200s each + overhead ≈ 30 min
+    /// 5 fan levels × up to 6 min each + overhead
     /// 120-second window (full time constant, ~95% accuracy)
     case optimized
 
     public var description: String {
         switch self {
-        case .quick: return "Quick (up to 15 min)"
-        case .standard: return "Standard (up to 22 min)"
-        case .optimized: return "Optimized (up to 30 min)"
+        case .quick: return "Quick (up to 17 min)"
+        case .standard: return "Standard (up to 25 min)"
+        case .optimized: return "Optimized (up to 35 min)"
         }
     }
 
@@ -274,8 +272,6 @@ public final class CalibrationRunner {
     }
 
     /// Performance ceiling — stop increasing load if temp reaches this
-    private static let performanceCeiling: Float = 85.0
-
     /// Target heating rate for calibration — matches real-world workloads.
     /// Research: real workloads heat Apple Silicon at ~1-2°C/sec.
     /// Max synthetic load heats at ~5-8°C/sec (Notebookcheck, Max Tech).
@@ -516,7 +512,6 @@ public final class CalibrationRunner {
         )
     }
 
-    /// Wait for machine to cool below a threshold
     /// Check if temperature readings have stabilized.
     /// Stable = stdev < 0.5°C AND slope < 0.05°C/sec over the window.
     private func isStabilized(readings: [Float]) -> Bool {

Sources/thermalforge/ThermalForge.swift

@@ -341,7 +341,7 @@ struct Calibrate: ParsableCommand {
         print("Mode: \(calMode.description)")
         print("Stress: \(calStress.description)")
         print("")
-        print("This will stress your \(calStress == .combined ? "CPU and GPU" : calStress == .cpu ? "CPU" : "GPU") and measure thermal response at 4 fan speed levels.")
+        print("This will stress your \(calStress == .combined ? "CPU and GPU" : calStress == .cpu ? "CPU" : "GPU") and measure thermal response at 5 fan speed levels.")
         print("Fans will be loud during the test.")
         print("")
         print("DISCLAIMER: Calibration pushes your Mac to full load and cycles fan speeds.")