fix: use sustainable native haptics cadence for mixed controllers

HAPTICS_EXPERIMENT.md

@@ -29,7 +29,7 @@ Sources:
 
 1. AIO/XInput defaults enable `SWITCH_PICO_HAPTICS_EXPERIMENT`, `SWITCH_PICO_HD_RUMBLE`, packet-level CYW43 reads and bounded HCI credit batching at 300 MHz/1.3 V. UART is unchanged. Preserve wake identity, pairing storage and USB modes. Incoming flow control and FIFO capacities remain unchanged; the controller's advertised outgoing capacity is eight ACL packets on this hardware.
 2. One selected Sony DualSense/DualSense Edge, Bluetooth Classic, sufficient negotiated MTU. Auto-arm chooses the first eligible ready controller, not necessarily slot 0, and later controllers do not steal an active stream. The fixture requires explicit start. Idle native output remains silent; other models and unselected slots use their controller-specific compatibility paths.
-3. Report 0x32 plus A2 remains a 143-byte L2CAP SDU. The first report selects native mode with sized state block 0x90/63 and one silent 0x92/64 haptic block. Subsequent reports use compact controls `{0x91,3,0x62,16,counter}`. Qualified gameplay carries one 64-byte block (32 stereo frames, descriptor 0x92); the fixture and unoptimized builds carry two blocks (64 frames, descriptor 0xd2). The counter advances by the number of blocks. Padding and Bluetooth CRC remain deterministic. No speaker, microphone, USB audio endpoint, Opus or resampler.
+3. Report 0x32 plus A2 remains a 143-byte L2CAP SDU. The first report selects native mode with sized state block 0x90/63 and one silent 0x92/64 haptic block. Subsequent reports use compact controls `{0x91,3,0x62,16,counter}`. Standard gameplay and the fixture carry two blocks (64 stereo frames, descriptor 0xd2). Explicit `SWITCH_PICO_HD_PACKET_FRAMES=32` carries one 64-byte block (descriptor 0x92) for single-controller qualification only. The counter advances by the number of blocks. Padding and Bluetooth CRC remain deterministic. No speaker, microphone, USB audio endpoint, Opus or resampler.
 4. At 3 kHz, 32/64 stereo frames require 93.75/46.875 reports/s. Absolute rational deadlines preserve fractional time and skip obsolete packets after stalls rather than burst-replaying them. Timer wakeups account for SDK +1 tick. Can-send permission and audio deadlines remain separate; flags are armed before requests and synchronous callbacks cannot recursively generate a stream.
 5. The deterministic fixture remains a finite 288-report / 6.144-second sequence: 48 priming intervals, four cycles of left 100 Hz / silence / right 200 Hz / silence (12 reports = 256 ms per phase), then 48 trailing-silence reports. Its peak remains 32/127. Gameplay is continuous, has no one-second priming pattern, and uses timestamped Switch commands instead. Stop restores compatibility output; disconnect cancels without stale-pointer use.
 6. No historical PCM FIFO. Generate only the current due block when transmission is permitted; bounded control mailbox across cores. Record packet counts, skipped blocks, failed sends, synchronous callbacks, generation cost, send gaps, lateness, request wait and first-tone timestamps. HCI submission is not physical actuator onset.
@@ -46,6 +46,7 @@ cmake -S . -B build-hd-rumble -DPICO_BOARD=pico2_w \
   -DSWITCH_PICO_HD_RUMBLE=ON \
   -DSWITCH_PICO_SYS_CLOCK_MHZ=300 -DSWITCH_PICO_OVERCLOCK_MV=1300 \
   -DSWITCH_PICO_CYW43_PACKET_READ=ON -DSWITCH_PICO_HCI_CREDIT_BATCH=ON \
+  -DSWITCH_PICO_HD_PACKET_FRAMES=64 \
   -DSWITCH_PICO_HAPTICS_EXPERIMENT_RAM=ON -DSWITCH_PICO_LOG=OFF
 cmake --build build-hd-rumble
 ```
@@ -56,7 +57,7 @@ The decoder preserves each actuator's one-to-three ordered substeps and frequenc
 
 The synthesizer has independent left/right low/high phase accumulators. Frequencies are `40 * 2^(index/32)` and `80 * 2^(index/32)` Hz. Each Switch command occupies an 8 ms window, split into 24/12/8 PCM samples per substep for counts 1/2/3. New reports supersede unplayed old substeps; identical compressed words hold final state rather than replaying deltas. Each Switch-updated side expires after 50 ms, matching the existing conservative timeout policy.
 
-One report interval of causal lookback is 10.667 ms in the qualified 32-frame path and 21.333 ms in the 64-frame path. Fixed 16-entry cross-core and synthesis histories contain commands, not PCM. Overflow is counted and obsolete sample intervals are not replayed. XInput holds use a distinct persistent command: strong/low magnitude drives the left 160 Hz band, weak/high drives the right 320 Hz band, until a new command or zero stop. They do not fake refreshes to evade the Switch watchdog. Retained XInput state is seeded once per native run, including manual re-arming after compatibility output.
+Standard gameplay uses 21.333 ms causal lookback; the explicit 32-frame experiment uses 10.667 ms. Fixed 16-entry cross-core and synthesis histories contain commands, not PCM. Overflow is counted and obsolete sample intervals are not replayed. XInput holds use a distinct persistent command: strong/low magnitude drives the left 160 Hz band, weak/high magnitude drives the right 320 Hz band, until a new command or zero stop. They do not fake refreshes to evade the 50 ms Switch watchdog. Retained XInput state is seeded once per native run, including manual re-arming after compatibility output.
 
 Native gameplay uses balanced **2x low/high gain after profile scaling**, followed by a gentle **0.8-power curve** on the combined amplitude. This lifts quiet and medium effects while retaining their low/high ratio. The curve is a 257-entry lookup with integer interpolation, not per-sample floating-point math. Combined weights are capped at 65535 to avoid overflow and clipping. Zero remains zero. The amplitude curve does not alter carrier frequencies or local-confirmation gain; packet timing follows the transport configuration above. This response replaced the initial 1.5x and low-band-only experiments after user comparison.
 
@@ -287,3 +288,26 @@ Pico/UART builds. The profile catalog was migrated and all sixteen stored
 profiles were compared with the pre-migration backup; the temporary editor
 profile and name were restored. No configuration, bond, or wake-identity reset
 was part of the transport work.
+
+### Mixed-controller cadence limit
+
+Final testing with a Switch Pro plus a DualSense and continuous USB motion
+reads changed the cadence decision. Rumble commands targeted only the
+DualSense; the Pro supplied concurrent input traffic, not a Pro rumble
+qualification. The 32-frame stream received all 2,050
+host commands but skipped **80 audio slots in 16.6 s**. Maximum permission
+wait reached **17,180 us**, exceeding its 10,667 us interval, with all eight
+outgoing credits observed in use. CPU clock remained 300 MHz/1.3 V.
+
+The standard build therefore uses **64 frames / 46.875 reports per second**,
+without reverting its CPU or transport improvements. The same 2,050-command
+mixed-controller comparison passed with zero drops/skips/send failures and
+778 audio reports. Its worst observed report gap was 26,588 us. A subsequent
+roughly 65-second mixed-controller stress run received all **8,194 commands**
+and submitted **3,082 audio reports**, with **zero drops, skipped audio slots,
+or send failures**. It processed 41,738 input reports and its worst observed
+audio report gap was 26,655 us.
+
+The 32-frame path remains an explicit single-controller experiment and is
+covered by the same native protocol/lifecycle tests; it is not advertised as
+sustainable for mixed/four-controller operation.