android: 修正增强档位说明（写反了）+ 分析页测温点沿用照片数值

上一提交把档位含义写成了「越高纳入越多弱对比区域」，方向是反的。按公式
amplitude = strength*(center-mean)/divisor, divisor = max(max-mean, mean-min, mapStrength),
mapStrength = strength*gain*2>>8 重新推导：
- 弱对比区域 divisor 被 mapStrength 兜住 → 幅度 = 128/gain，与档位无关；
- 强边缘 divisor 由对比度主导 → 幅度随档位线性增长；
- 门槛 mapStrength <= range*32 → 档位越高弱对比区域被整个排除（正是噪声所在处）。
即高档位是「更锐但更挑」。设置页文案与 DetailEnhance 文档已改正，
并加 theLevelRaisesTheContrastGateAndDropsWeakDetail 锁定
（range=60 窗口：8 档 mapStrength 1500 通过、16 档 3000 被拒）。
真机 HF 实测 0.365(关) / 0.438(6档) / 0.433(8档)，与推导一致。

分析页测温点原先前先重新测量 NUC，于是照片上烧录的 Pt2 26.1℃ 与面板 25.9℃
不一致（与极值同一类漂移）。现在优先用文件记录的 tempMc——照片上印的就是它，
仅在记录缺失时才测量。真机确认照片/面板/极值三方数值完全一致。

真机（小米 22041211AC/Android12/MIUI，重启后冷启动）：15.1fps 稳定、单帧 6.1~6.6ms、
6/8 档 9.9~10.3ms、拍照 extremes=2、分析页只有一个 min/max 且与照片重合、
底部面板可见（36.0/20.4/24.5）。101 项测试全绿。

android/app/src/main/kotlin/com/mag160c/thermal/core/DetailEnhance.kt

@@ -44,6 +44,18 @@ package com.mag160c.thermal.core
  * near-silent on flat areas). The divisor floor is what keeps flat, noisy regions
  * quiet while edges still get pushed.
  *
+ * ## What the strength level actually controls
+ *
+ * Raising the level does two things at once:
+ *   1. it RAISES the contrast gate — a window is enhanced only when its 7x7 range
+ *      clears `mapStrength / 32`, so low-contrast areas (where sensor noise lives)
+ *      are excluded entirely at high levels; and
+ *   2. it RAISES the amplitude on strong edges, where the local contrast itself
+ *      dominates the divisor.
+ * Where the contrast is small enough that the divisor floor applies, the amplitude
+ * works out to `strength / mapStrength` = `128 / gain` — independent of the level.
+ * That is why turning the level up sharpens the picture instead of boiling it.
+ *
  * ## Validation status — READ THIS
  *
  * The arithmetic above is a faithful transcription, and the structural properties

android/app/src/main/kotlin/com/mag160c/thermal/ui/analyze/AnalyzeViewModel.kt

@@ -102,8 +102,15 @@ class AnalyzeViewModel(
             // which showed up on device as two "min" markers a few pixels apart.
             if (recorded.minPos >= 0) mnPos = recorded.minPos
             if (recorded.maxPos >= 0) mxPos = recorded.maxPos
+            // Each probe keeps the temperature the CAPTURE recorded, because that is
+            // the number burned into the photo next to the marker. Re-measuring the
+            // NUC block gives a slightly different value (the sensor drifts between
+            // capture and reload), which showed on device as the photo saying
+            // "Pt2 26.1" while the panel said 25.9 for the same point. Only probes
+            // that somehow lack a recorded value are measured.
             val loaded = (parsed?.probes.orEmpty()).map { p ->
-                Probe(p.x, p.y, p.label, measureSensor(p.x, p.y) ?: (p.tempMc / 1000f))
+                Probe(p.x, p.y, p.label, (p.tempMc / 1000f).takeIf { p.tempMc != 0 }
+                    ?: measureSensor(p.x, p.y) ?: 0f)
             }
             withContext(Dispatchers.Main) {
                 _render.value = bmp

android/app/src/main/kotlin/com/mag160c/thermal/ui/settings/SettingsScreen.kt

@@ -247,17 +247,21 @@ fun SettingsScreen(
             text = {
                 Column {
                     Text(
-                        "加强局部细节（官方同款 7×7 局部映射）。档位沿用官方 SDK 自己的" +
+                        "加强局部细节（官方同款 7×7 局部映射）。档位越高：①对比度门槛越高，" +
-                            "定义：档位越高，纳入增强的弱对比区域越多（边缘强度本身不变），" +
+                            "只有更强的细节会被增强，弱对比区域（噪声所在处）被排除；" +
-                            "画面更通透，弱细节里的噪声也会一起显现。",
+                            "②强边缘的增强幅度更大。所以高档位是「更锐但更挑」，" +
+                            "而不是「连噪声一起放大」。",
                         style = MaterialTheme.typography.bodySmall,
                         color = MaterialTheme.colorScheme.onSurfaceVariant,
                         modifier = Modifier.padding(bottom = 8.dp),
                     )
                     // Vendor levels: MAG_SetDetailEnhancement accepts 0..32 and the
-                    // pipeline uses `level << 3`. Extra headroom is safe because the
+                    // pipeline uses `level << 3`. The level raises BOTH the contrast
-                    // per-edge amplitude is level-independent (128/gain); the level
+                    // gate (a window qualifies only when its 7x7 range clears
-                    // only decides which windows clear the contrast threshold.
+                    // mapStrength/32) and the amplitude on strong edges (where the
+                    // local contrast dominates the divisor). On flat, noisy areas the
+                    // divisor floor holds the amplitude at 128/gain regardless of
+                    // level, which is why cranking it up sharpens instead of boiling.
                     Column(
                         modifier = Modifier
                             .heightIn(max = 320.dp)

android/app/src/test/kotlin/com/mag160c/thermal/core/DetailEnhanceTest.kt

@@ -184,6 +184,53 @@ class DetailEnhanceTest {
         )
     }
 
+    /**
+     * What the level knob really does (this is the sentence the settings dialog
+     * shows the user, so it is pinned here):
+     *   - the level raises the contrast GATE: a window is enhanced only when its 7x7
+     *     range clears `mapStrength / 32` (= level * gain * 2 / 256 / 32), so weak
+     *     detail — where sensor noise lives — is dropped at high levels; and
+     *   - where a window IS gated in, the divisor floor keeps its amplitude near
+     *     `128/gain` of the deviation, so it does not grow with the level.
+     *
+     * Window used below: 15 samples at 7200 and one at 7260 (range 60), centre 7260,
+     * gain 3000.
+     *   level 8  (strength 64 -> mapStrength 1500): gate 1500 <= 60*32 = 1920 passes
+     *   level 16 (strength 128 -> mapStrength 3000): gate 3000 <= 1920 FAILS -> dropped
+     */
+    @Test
+    fun theLevelRaisesTheContrastGateAndDropsWeakDetail() {
+        val size = 16
+        fun window(): IntArray {
+            val src = IntArray(size * size)
+            for (dy in 0 until 4) for (dx in 0 until 4) src[(dy * 2) * size + dx * 2] = 7200
+            src[0] = 7260
+            src[3 * size + 3] = 7260
+            return src
+        }
+        fun changeAt(strength: Int): Int {
+            val gray = ByteArray(size * size) { 128.toByte() }
+            DetailEnhance(size, size).enhance(window(), gray, strength = strength, gain = 3000)
+            return (gray[3 * size + 3].toInt() and 0xFF) - 128
+        }
+        assertTrue(
+            "level 8 (mapStrength 1500) must act on a range-60 window",
+            changeAt(64) != 0,
+        )
+        assertEquals(
+            "level 16 (mapStrength 3000) must drop the same window: its contrast " +
+                "is below the gate",
+            0,
+            changeAt(128),
+        )
+        // and the amplitude is small wherever it does apply: the divisor floor caps
+        // it near 128/gain, so higher levels cannot make flat areas boil
+        assertTrue(
+            "amplitude must stay small, was ${changeAt(64)}",
+            kotlin.math.abs(changeAt(64)) <= 4,
+        )
+    }
+
     @Test
     fun doesNotChangeTheOverallLevelOfASmoothImage() {
         // on a smooth ramp the mean and the centre are equal, so no delta should be

docs/android_app/session_state.md

@@ -578,11 +578,24 @@ extremes **合并成一次 draw**（原来分两次调用，彼此看不见）
 SetDetailEnhancement**（jadx 全量搜索无此符号），所以官方实时画面用的是
 SDK 内部默认档——无法据此对齐具体档位，故本应用把它做成用户可见设置
 （0/1/2/3/4/6/8 档），默认**关闭**（关闭时 `RenderPipelineTest` 逐位基线不受影响）。
-一个有用的推论：每档的边缘强度其实与档位无关（化简为 `128/gain`），
-档位只决定哪些窗口通过对比度阈值——所以调高档位是"纳入更多弱对比区域"，
-而不是"把边缘推得更狠"。
 
-实机实测：级别 2 下仍 15.1fps，单帧绘制 11.3ms（滤波器约 +5ms，帧预算 66ms 内）。
+**档位到底控制什么（推导 + 真机 HF 实测，曾一度写反）**：
+`amplitude = strength*(center-mean)/divisor`，`divisor = max(max-mean, mean-min, mapStrength)`，
+`mapStrength = strength*gain*2>>8`（gain=1000 → 每档 62.5）。由此：
+- 弱对比区域：divisor 被 mapStrength 兜住 → 幅度化简为 `128/gain`，**与档位无关**；
+- 强边缘：divisor 由局部对比度主导 → 幅度**随档位线性增长**；
+- 门槛 `mapStrength <= range*32` → 档位越高，**弱对比区域被整个排除**（噪声所在处）。
+所以高档位是"更锐但更挑"，不是"连噪声一起放大"（设置页文案曾写反，已改）。
+真机 HF 能量实测：关 0.365、6 档 0.438、8 档 0.433（6/8 相近，符合"8 档门槛更高、
+只增强更强边缘"）。该行为由 `theLevelRaisesTheContrastGateAndDropsWeakDetail` 锁定
+（window range=60：8 档 mapStrength 1500 通过、16 档 3000 被拒）。
+
+**另一个一致性修复**：分析页原先把测温点拿去**重新测量** NUC，导致照片上烧录的
+`Pt2 26.1℃` 与面板显示的 25.9℃ 不一致（同极值那类漂移）。现在优先采用文件里
+记录的 `tempMc`（照片上印的就是它），仅在缺失时才测量。
+
+实机实测：级别 2 下仍 15.1fps，单帧绘制 11.3ms（滤波器约 +5ms，帧预算 66ms 内）；
+6/8 档 9.9~10.3ms，同样不掉帧。
 
 ### 8) 其他顺带修复（本轮真机复现）