docs: add HANDOFF.md — cross-machine pickup for the audio work

Captures branch map, the SpkTrim sweep-test protocol, open threads (#10/#11/defaults soak-test), why DS5 BT audio is hard, gotchas, and a one-paste kickoff for a fresh session. Written so the main-PC session can resume from git alone.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

CHANGELOG.md

@@ -8,6 +8,10 @@ Format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/). Version
 
 ## [Unreleased]
 
+### Added (experimental — `audio/speaker-rate-trim` branch)
+
+- **`SpkTrim` setting — fine speaker clock-drift trim to chase the periodic crackle (#7).** The DS5's DAC and the USB host's audio clock are independent ~48 kHz crystals; the residual ppm mismatch slowly drifts the controller's audio buffer, which underruns and crackles. The speaker path now delivers `48000 + trim` samples/s (via a zero-order-hold duplicate/drop accumulator) instead of exactly 48000, to null that drift. New `speaker_rate_trim` config field (stored 0–200 = **−100…+100 Hz**, default **0 Hz** = exact no-op), swept on the OLED **Settings** screen with the D-pad ▶◀ at 1 Hz/step. **Experimental / diagnostic:** if one value silences the crackle and *holds*, the drift was a static offset and this is the fix; if it nulls then creeps back, that proves the drift is dynamic and needs adaptive resampling. Tune with `scripts/sine_ch12.py` playing a steady tone and time the interval between crackles.
+
 ---
 
 ## [0.6.11-oled-edition] — 2026-06-02

HANDOFF.md

@@ -0,0 +1,106 @@
+# HANDOFF — audio work in progress
+
+_Last updated: 2026-06-04 (written on the laptop; continuing on the main PC)_
+
+This is a working handoff so the next session — on the main PC, or a fresh Claude Code
+session — can pick up exactly where we left off. **All the real work is committed and pushed
+to `origin`; nothing is stuck on the laptop.** This file lives on the `audio/speaker-rate-trim`
+branch, so after `git checkout audio/speaker-rate-trim` it's right here.
+
+---
+
+## TL;DR
+
+- The frustration that kicked this off: **audio (speaker/mic/HD haptics) is the priority**, not OLED features. Goal: make the dongle behave **as close to a wired-USB DualSense as possible**, with flaky extras opt-in.
+- We confirmed the open GitHub issues are real, synced the tree up to **v0.6.11** (it was stale at 0.6.4), and split the work onto two feature branches.
+- **Main thing to do next:** build `audio/speaker-rate-trim` on the main PC, flash it, and run the **`SpkTrim` sweep test** (below) to find out whether the speaker crackle is a fixed clock offset (curable with the new knob) or live drift (needs adaptive resampling).
+
+---
+
+## Branch map
+
+| Branch | Commit | State | Needs |
+|---|---|---|---|
+| `master` | `878a742` | untouched, = `origin/master` = **v0.6.11** | — |
+| `defaults/usb-faithful` | `333201e` | pushed | **HIL soak-test** before merging to master |
+| `audio/speaker-rate-trim` | `3b157cb` | pushed | **build + flash + sweep test** (not yet built) |
+
+Both feature branches are pushed to `origin` (the fork) and tracking. `upstream` / `upstream-fork` were **not** touched.
+
+---
+
+## What each branch contains
+
+### `defaults/usb-faithful` — out-of-box "acts like USB" defaults
+A full audit found the firmware is already ~USB-faithful for everything that matters (input is byte-for-byte passthrough; lightbar defaults to host control; etc.). Only two defaults were flipped:
+- **`polling_rate_mode` 0 → 2** (250 Hz → **1000 Hz / realtime**) — matches a wired DS5's 1 ms latency. ⚠️ Realtime also drops the report-throttle, so **this needs a 30–60 min hardware soak-test** (watch for BT drops / OLED Diag `BT31 in/s` collapse) before it merges to master. Fallback if unstable: `polling_rate_mode = 1` (500 Hz).
+- **`bt_mic_enable` 1 → 0** (mic **off** by default) — the BT mic has a known 2× playback bug (#10), so it's opt-in until fixed.
+
+Both only affect fresh flashes / Reset-to-defaults; existing saved configs keep their values.
+
+### `audio/speaker-rate-trim` — experimental crackle fix (the active thread)
+New **`SpkTrim`** setting that trims the speaker sample rate to null clock drift.
+
+**Why:** the speaker path is 1:1 (no resampler) — it delivers samples at the **USB host's 48 kHz clock** while the DS5 consumes at **its own 48 kHz crystal**. Two independent clocks drift apart → the DS5's audio buffer slowly underruns → the periodic **crackle (#7)**. 0.6.11's "retiming" fixed the *average* rate but can't track *drift*.
+
+**What it does:** delivers `48000 + trim` samples/s via a zero-order-hold duplicate/drop accumulator (`src/audio.cpp`, in the speaker accumulation loop). At ppm scale, a single duplicated/dropped sample is inaudible vs a 480-sample gap.
+
+**Files changed:** `src/config.h` (field), `src/config.cpp` (clamp), `src/audio.cpp` (the trim), `src/oled.cpp` (Settings UI), `CHANGELOG.md`.
+
+**Config field:** `speaker_rate_trim`, stored `0..200` = **−100…+100 Hz**, default **100 (= 0 Hz, an exact no-op)**. So a fresh flash sounds **identical to 0.6.11** until you sweep it — safe to flash.
+
+**UI:** OLED **Settings** screen, item **`SpkTrim ±NHz`** (just above "Reset to defaults"), swept with the D-pad **▶ / ◀** at **1 Hz/step**. (Reset/Wipe moved to settings indices 16/17 via named constants.)
+
+---
+
+## Next steps on the main PC
+
+```bash
+git fetch origin
+git checkout audio/speaker-rate-trim
+
+# ensure TinyUSB is still pinned (CLAUDE.md — 0.18 that ships with SDK 2.2.0 won't compile):
+( cd "$PICO_SDK_PATH/lib/tinyusb" && git checkout 0.20.0 )
+
+cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DPICO_SDK_PATH="$PICO_SDK_PATH"
+cmake --build build --target ds5-bridge        # → build/ds5-bridge-oled.uf2
+# BOOTSEL the board, copy build/ds5-bridge-oled.uf2 onto the RP2350 mount
+```
+
+### The sweep test (the whole point)
+1. OLED **Settings** → scroll to **`SpkTrim`** (reads `+0Hz`).
+2. Play a steady tone from the host: `python3 scripts/sine_ch12.py 600` (10 min of 440 Hz on the speaker channels).
+3. Nudge `SpkTrim` with the D-pad ▶/◀ and **time the interval between crackles**. A longer interval = closer to the DS5's true rate. Walk it toward the value where the crackle stops. (The crackle interval is itself a tachometer: a click every ~16 s ≈ ~600 ppm of clock error.)
+4. **Decision read — leave the good value running 10–15 min:**
+   - **Stays clean → static offset.** The knob is the fix; we lock in a default and ship it.
+   - **Creeps back → dynamic drift.** Proof we need **adaptive resampling** (continuously measure drift and vary the trim), which is the bigger but "correct" fix. The DS5 gives no buffer back-channel, so by-ear / crackle-interval is the right instrument.
+
+---
+
+## Other open threads (not blocking the sweep)
+
+- **`defaults/usb-faithful` soak-test** — validate 1000 Hz polling stability, then merge to master (or fall back to 500 Hz).
+- **#10 mic 2× playback** — needs OLED **Diagnostics** `Mic dec=` (samples per `opus_decode`; `240` would confirm a half-rate stream) and `Mic in/s` (arrival rate) read with the mic enabled. Mic is off-by-default now, so it's no longer an out-of-box problem.
+- **#11 constant low-level haptics** — root-caused to the **#6 adaptive-trigger fix** (`71cead4`): 0.6.11 now honors host trigger-FFB, and the dongle re-asserts `state[]` continuously via the `0x36` audio frames, so a game's idle trigger effect buzzes constantly. **Not** auto-haptics (already off). Decisive read: watch the OLED Diag **`trig`** / **`host02`** counters during the buzz — climbing = host streaming FFB (add an opt-out toggle); static = stale latch (clear FFB params when allow-bit is 0, in `src/state_mgr.cpp`).
+
+---
+
+## Why audio is hard (context for any new session)
+
+The DualSense's Bluetooth audio is **not A2DP / not normal Bluetooth audio** — Sony tunnels Opus-encoded audio inside the HID controller reports (`0x31`/`0x36`) over a proprietary, undocumented protocol with no clock negotiation. A normal PC/phone paired to a DS5 over BT gets **zero** controller audio; **only the PS5** implements it (Sony built both ends + a dedicated radio). This dongle reverse-engineers that on a CYW43 chip over emulated SPI (hence the load-bearing 320 MHz overclock) with general-purpose BTstack. So the gap from PS5-quality is **fidelity, not feasibility** — the crackle and mic-2× are timing/rate bugs, not walls.
+
+---
+
+## Gotchas / reminders
+
+- **Push only to `origin`** — never `upstream` / `upstream-fork`.
+- **TinyUSB must be 0.20.0** or the build fails on `TUD_AUDIO_EP_SIZE`.
+- Build target is `ds5-bridge`; output is `build/ds5-bridge-oled.uf2` (custom name).
+- `SpkTrim = 0` is a byte-exact no-op, so flashing this branch can't regress audio.
+- The `0x36` packet layout and the `state[]` re-assertion are **load-bearing** (see CLAUDE.md) — don't "simplify" the audio frame.
+
+---
+
+## One-paste kickoff for a fresh Claude session
+
+> I'm on branch `audio/speaker-rate-trim`. Read `HANDOFF.md`, the latest commit message, and the `[Unreleased]` section of `CHANGELOG.md`. We added an experimental `SpkTrim` setting to chase the speaker crackle (#7) by trimming the speaker sample rate to null clock drift between the host and the DualSense. Help me build the UF2, flash it, and run the sweep test in HANDOFF.md — find the `SpkTrim` value where the crackle stops, then check whether it holds (static offset = fix) or creeps back (dynamic drift = needs adaptive resampling).

src/audio.cpp

@@ -264,12 +264,20 @@ void audio_loop() {
 
     static float audio_buf[480 * 2];
     static uint audio_buf_pos = 0;
+    static double spk_acc = 0.0;   // speaker rate-trim accumulator (persists across calls)
     // 2. 从4ch中提取ch3/ch4，转换为float输入重采样器
     WDL_ResampleSample *in_buf;
     int nframes = resampler.ResamplePrepare(frames, OUTPUT_CHANNELS, &in_buf);
 
     const float audio_gain = mute[0] ? 0.0f : powf(10.0f, get_config().speaker_volume / 20.0f);
     const float haptics_gain = get_config().haptics_gain;
+
+    // EXPERIMENTAL speaker clock-drift trim (audio/speaker-rate-trim). Deliver
+    // 48000 + trim samples/s to the DS5 instead of exactly 48000, to null the slow
+    // inter-crystal drift behind the crackle (#7). trim = 0 → ratio == 1.0 → exact
+    // no-op (byte-identical to pre-trim, since 48000/48000 is exact in double).
+    const int    spk_trim_hz = (int)get_config().speaker_rate_trim - 100; // -100..+100 Hz
+    const double spk_ratio   = (48000.0 + spk_trim_hz) / 48000.0;
     uint16_t spk_max = g_peak_spk;
     uint16_t hap_max = g_peak_hap;
 
@@ -308,19 +316,32 @@ void audio_loop() {
             if (b > hap_max) hap_max = b;
         }
  #if !DISABLE_SPEAKER_PROC
-        audio_buf[audio_buf_pos++] = raw[i * INPUT_CHANNELS] / 32768.0f * audio_gain;
+        {
-        audio_buf[audio_buf_pos++] = raw[i * INPUT_CHANNELS + 1] / 32768.0f * audio_gain;
+            const float spk_l = raw[i * INPUT_CHANNELS]     / 32768.0f * audio_gain;
-        if (audio_buf_pos == 480 * 2) {
+            const float spk_r = raw[i * INPUT_CHANNELS + 1] / 32768.0f * audio_gain;
-            static audio_raw_element element{};
+            // Rate-trim: emit `spk_ratio` output samples per input sample on average
-            memcpy(element.data, audio_buf, 480 * 2 * 4);
+            // (usually 1; occasionally 2 when trim>0, or 0 when trim<0). Zero-order
-            if (queue_is_full(&audio_fifo)) {
+            // hold — at ppm-scale trims a single duplicated/dropped sample is inaudible
-                queue_try_remove(&audio_fifo,NULL);
+            // vs a full 480-sample buffer underrun.
-                g_fifo_drops++;
+            spk_acc += spk_ratio;
+            int emit = (int)spk_acc;
+            spk_acc -= emit;
+            while (emit-- > 0) {
+                audio_buf[audio_buf_pos++] = spk_l;
+                audio_buf[audio_buf_pos++] = spk_r;
+                if (audio_buf_pos == 480 * 2) {
+                    static audio_raw_element element{};
+                    memcpy(element.data, audio_buf, 480 * 2 * 4);
+                    if (queue_is_full(&audio_fifo)) {
+                        queue_try_remove(&audio_fifo,NULL);
+                        g_fifo_drops++;
+                    }
+                    if (!queue_try_add(&audio_fifo, &element)) {
+                        printf("[Audio] Warning: audio_fifo add failed\n");
+                    }
+                    audio_buf_pos = 0;
+                }
             }
-            if (!queue_try_add(&audio_fifo, &element)) {
-                printf("[Audio] Warning: audio_fifo add failed\n");
-            }
-            audio_buf_pos = 0;
         }
 #endif
         float h_l = raw[i * INPUT_CHANNELS + 2] / 32768.0f * haptics_gain;

src/config.cpp

@@ -121,6 +121,10 @@ void config_valid() {
         body->controller_wakes_display = 1;
         printf("[Config] controller_wakes_display invalid, defaulting to 1 (on)\n");
     }
+    if (body->speaker_rate_trim > 200) {  // 0xFF erased / upgrade → 0 Hz (centered)
+        body->speaker_rate_trim = 100;    // 100 encodes 0 Hz; actual trim = stored - 100
+        printf("[Config] speaker_rate_trim invalid, defaulting to 100 (0 Hz)\n");
+    }
     if (body->config_version != CONFIG_VERSION) {
         body->config_version = CONFIG_VERSION;
         printf("[Config] Warning: Config may breaking change\n");

src/config.h

@@ -54,6 +54,14 @@ struct __attribute__((packed)) Config_body {
     // preserves the original "any controller activity wakes the screen"
     // behavior. Issues #8 (dim timeout never fired during play) and #9.
     uint8_t controller_wakes_display;
+    // EXPERIMENTAL (audio/speaker-rate-trim branch): fine speaker clock-drift trim.
+    // The DS5's DAC and the USB host's audio clock are independent ~48 kHz crystals;
+    // the residual ppm mismatch slowly drifts the DS5's buffer and produces the
+    // periodic crackle (#7). This nudges our delivered rate to 48000 + (trim-100) Hz
+    // to null that drift. Stored 0..200, offset by 100 so an erased-flash 0xFF clamps
+    // to 100 = 0 Hz (an exact no-op, byte-identical to pre-trim). Swept on the OLED
+    // Settings screen with the D-pad ▶◀, 1 Hz/step. Default 100 (0 Hz).
+    uint8_t speaker_rate_trim;
 };
 
 struct __attribute__((packed)) Config {

src/oled.cpp

@@ -130,7 +130,7 @@ constexpr int kLbModeHost = 8;
 constexpr int kNumLbModes = 9;
 
 // Settings screen state
-constexpr int kNumSettingsItems = 17; // 8 fields + 3 auto-haptic + 2 screen-timeout + BT mic + Ctrl-wake + Reset + Wipe
+constexpr int kNumSettingsItems = 18; // 8 fields + 3 auto-haptic + 2 screen-timeout + BT mic + Ctrl-wake + Spk-trim + Reset + Wipe
 constexpr int kSettingsAutoHapEnaIdx  = 8;
 constexpr int kSettingsAutoHapGainIdx = 9;
 constexpr int kSettingsAutoHapLpIdx   = 10;
@@ -138,8 +138,9 @@ constexpr int kSettingsScrDimIdx      = 11;
 constexpr int kSettingsScrOffIdx      = 12;
 constexpr int kSettingsBtMicIdx       = 13;
 constexpr int kSettingsCtrlWakeIdx    = 14;
-constexpr int kSettingsResetIdx       = 15;
+constexpr int kSettingsSpkTrimIdx     = 15;
-constexpr int kSettingsWipeSlotsIdx   = 16;
+constexpr int kSettingsResetIdx       = 16;
+constexpr int kSettingsWipeSlotsIdx   = 17;
 Config_body settings_local{};
 int settings_sel = 0;
 bool settings_dirty = false;
@@ -1599,6 +1600,12 @@ void settings_adjust(int delta) {
         }
         case 13: c.bt_mic_enable ^= 1; break; // BT mic on/off
         case 14: c.controller_wakes_display ^= 1; break; // controller activity wakes OLED on/off
+        case 15: { // speaker_rate_trim  stored 0..200 = -100..+100 Hz, 1 Hz/step
+            int v = (int)c.speaker_rate_trim + delta;
+            if (v < 0) v = 0; if (v > 200) v = 200;
+            c.speaker_rate_trim = (uint8_t)v;
+            break;
+        }
     }
 }
 
@@ -1702,8 +1709,9 @@ __attribute__((noinline)) void format_settings_item(int idx, char* line, size_t
             break;
         case 13: snprintf(line, n, "%s BT Mic %s", cur, c.bt_mic_enable ? "on" : "off"); break;
         case 14: snprintf(line, n, "%s CtrlWake %s", cur, c.controller_wakes_display ? "on" : "off"); break;
-        case 15: snprintf(line, n, "%s Reset to defaults", cur); break;
+        case 15: snprintf(line, n, "%s SpkTrim %+dHz", cur, (int)c.speaker_rate_trim - 100); break;
-        case 16: snprintf(line, n, "%s Wipe all slots", cur); break;
+        case 16: snprintf(line, n, "%s Reset to defaults", cur); break;
+        case 17: snprintf(line, n, "%s Wipe all slots", cur); break;
     }
 }