POWER-FINAL-17 — modem request gate and screen-first wake

Status: INSTALLED, FIRST BOOT AND FIRST SLEEP CLEAN; tagged baseline-17 (2026-10-13). Package out/y2linux-platform-v1-power-final-candidate-17/ (fallback = 16), receipts out/deep-sleep-final/build17-receipts/. BOOTIMG 2ddd8d384cf6d9336d1a3505944f6a787764e982f7327400b5e886a41ea52d9a, Y2ROOT dc1f057a6c1d5768cbe3822893b5f71d3c3615bfbb3848e78491d9f0882a3e76. Flash BOOTIMG + ANDROID only with MT6582_preserve_data_scatter.txt, Download Only. Identity: build Y2LINUX-PLATFORM-V1-POWER-FINAL-17, kernel 6.18.56-y2linux-platform-v1-power-final-17, root 2026.08-platform-v1.47, release 1.0.0-platform-v1-power-final-candidate.17; Linux 0459116 on main (later: a test-only commit), Reborn unchanged (217e28d). Base: Baseline 16 (16 report). Scope audit: "POWER-FINAL-17 sleep-floor and wake-latency scope audit" in the roadmap gap audit.

Goal set by the owner: stock sleep parity, ≈0.1 points/h (≈1.3 mA), with deep suspend and resume working every time, and a faster wake (screen first, radios afterwards). 16's first night drew ≈8.6 mA.

What 17 changes

Area	Change	Switch
Modem (MD1) clock requests	The never-used modem's 26 peripheral clock requests are gated (PERICFG MD1_SET 0x20, readback MD1_STA 0x30) once the UI is up, and re-checked at every sleep entry. Stock golden has them all gated; 16 read 0 at all 33 sleep entries. Proven AP-writable and harmless by the live test on 16 (ledger F11). Main suspect for INFRA never powering down.	`md1_gate`
Wake order	After a wake by the owner, the sleep helper asks Reborn to light the screen first and restores Wi-Fi and Bluetooth afterwards. On 16 the screen waited ≈1.9 s for the radios' firmware and calibration (archive 110: SPM return 217.92 s, screen ≈220.45 s). Expected: screen ≈1 s after Power.	— (a failed early restore is retried at completion, as before)
From 16's follow-ups (`d1953fa`)	The MMSYS part of the stock DCM now reaches the image; the USB DMA-unmask readback uses bits 15..8 (no more false `dma_unmask_lost`).	—

md1_gate lives in /data/system/platform/power-features.json with the 16 switches, e.g. {"md1_gate": false}; the SPM state shows md1_gate=1 md1_result=0 md1_sta=0x3ffffff md1_regates=N.

Tests

Locked-container production suite: 682 cases, exit 0 (16: 676; new: MD1 gate policy/wiring/switch, wake order, early restore reuse). The overlay-hash test added after 16 failed on its first container run because the container's /src already carries the overlays (the kernel was fine: it contains the 0011 MMSYS DCM hook); the test now reverses each patch first (ebdfcf8). Package validation, ELF closure (354 files), installed ARM (40 modules), privacy and inventory PASS.

Hardware results

First boot (2026-10-13, boot de37f6bc, archive out/deep-sleep-final/evidence/20261013/111-power17-first-check/): kernel -power-final-17, build 0459116, taint 0. First frame at 24 s, application_ready at 44.7 s (16: 40.9 s). md1_gate=1 md1_result=0 md1_sta=0x3ffffff (dmesg Y2SPM: MD1 requests gated result=0), dcm_result=0, golden_apply=0xf, dis_pdn=1. Wi-Fi (wlan0) and Bluetooth (hci0) up, USB gadget configured (stage 6) and SSH working, charging from a PC port. Known and unchanged from 16: reborn-splash /dev/kmsg EINVAL at 25 s, mmc_erase group start error, thermal_zone1 -61. Not yet seen: a sleep, the wake time, audio.

First sleep (2026-10-13, archive out/deep-sleep-final/evidence/20261013/112-power17-short-sleep/): owner reports audio works and the wake feels faster. Unplugged sleep 11:08:39 → Power 11:26:32 (18 min), 1 entry / 1 resume / 0 aborts, wake 0x20; USB worked after the replug.

Wake: screen ≈0.5 s after the SPM return (SPM 373.19 s → DSI modeset 373.70–373.72 s; radios after it, WMT
374.06–375.40 s, wlan0 376.41 s). 16: ≈2.5 s (archive 110).
MD1 gated at entry: golden PERI_PDN0_MD1_STA e=3ffffff 1, md1_regates=0; SoC golden differences 11 → 10,
PMIC 15/44 unchanged.
INFRA still never powers down: last_lost=0x0, pdn0 0x2fcffff unchanged. The modem requests were not the (only)
blocker; next: SRCLKENA, PCM R13 0xe8060000 bits, MD_MASK 0x38 = 3.
USB unmask really lost once: with the readback fixed, dma_unmask_rearms 1 → 4 and dma_unmask_lost 0 → 1
across the unplug/sleep/restore (journal 11:03 vs 11:26). The re-arm restored it and USB worked: first direct support for the Defect B theory (which of the three re-arms found it lost is not recorded).
Two automatic sleep requests right after the unplug (11:08:12, 11:08:20) were refused after the radio quiesce
(complete_result_2); the third (11:08:36) entered. The refusal reason is not persisted (journal rows have no reason field and suspend-last.json is overwritten): diagnostics gap for 18.
18 min is too short for a drain figure.

What to check after the flash

cat /sys/devices/platform/10006000.power-controller/state: md1_gate=1 md1_result=0 md1_sta=0x3ffffff; USB,
Wi-Fi, Bluetooth, charger and audio as on 16.
Wake: press Power after a sleep and judge the time to the screen; read restoring_radios vs the DSI modeset in
dmesg (screen now before the radios).
After one sleep: golden row PERI_PDN0_MD1_STA now matches (e=3ffffff 1), and y2_infra/state last_lost
(non-zero = INFRA finally powers down).
Overnight drain against 16's ≈8.6 mA.