POWER-FINAL-27 — stock sleep-parity gaps C, A, B

Status: SEALED, AWAITING THE OWNER FLASH (2026-10-14). Package out/y2linux-platform-v1-power-final-candidate-27/ (fallback = 26), receipts out/deep-sleep-final/build27-receipts/. BOOTIMG c9017de1d19924c404f0776b9bacd0e09df9bca1c82802fc7f260e058c57902c, Y2ROOT c70dc92bb86a883ef83a0af22bbda52f6cd19577710f15a8532d576fa73f737a. Identity: build Y2LINUX-PLATFORM-V1-POWER-FINAL-27, kernel 6.18.56-y2linux-platform-v1-power-final-27, root 2026.08-platform-v1.57, release 1.0.0-platform-v1-power-final-candidate.27; Linux 2c3e2bd (changes f1d77a0, identity/audit df15aab, include fix 2c3e2bd), Reborn 1c9edd4 (unchanged).

Why

26's night (archive 20261014/124): 10.3 h, 21/21 clean sleeps, whole-night ≈8.3 mA, but the voltage slope fell from −10.6 to −1.2 mV/h through the night (post-charge relaxation), so the late-night drain is only bounded to ≈2–8 mA. The read-only stock comparison (ledger F18–F27, docs/knowledge/stock-suspend-re-ledger.md) found four real differences; 27 implements the three without storage risk.

What 27 changes

Gap	Change	Switch (`/data/system/platform/power-features.json`)	Kernel default
C panel	Screen-off sends the stock GC9503V `lcm_suspend` table (vendor pre-sequence, display off, sleep-in, 25 analog writes) and leaves reset high 50 ms (`panel.c`). Any DCS failure falls back to 26's display-off, sleep-in, reset-low. Screen-on unchanged: reset pulse + full init.	`panel_stock_suspend` (`/sys/module/panel/parameters/stock_suspend`)	N (26)
A VGP2	DTS `regulator-always-on` removed from VGP2 (CS43131 VA/VP/VCP, its only user in stock). The audio card holds one reference unless the switch is set; then the codec's runtime PM turns VGP2 off with VD/VL while idle (GPIOs first, VGP2 last, as stock `cs43131_power_down_hw`). DTB validator requires the new contract.	`vgp2_idle_off` (`/sys/module/y2_cs43131/parameters/vgp2_idle_off`)	N (hold, 26)
B VSYS	`pmic_sleep` group 8: stock PMIC_INIT VSYS fields (slew rates, `TRANSTD` 3, `VOSEL_TRANS_EN`, `VSLEEP_EN`, `QI_VSYS_VSLEEP` 2, `VOSEL_CTRL` 1 last) inside the SPM enter, restored exactly (mode bit first) before any resume. Only while VSYS is in software mode with VOSEL = VOSEL_ON = VOSEL_SLEEP = readback (live 0x40, 2.2 V), else refused and counted (`vsys_refusals`). VSYS selectors stay unwritable: the voltage never changes.	`pmic_sleep` (15 = with VSYS, 7 = 26)	0

power_features arms all three at application_ready; the sleep diagnostics rows record panel_stock_suspend, vgp2_idle_off and pmic_sleep_state (now with vsys_refusals).

Not changed: gap D (VEMC3V3, eMMC power-off needs a new storage suspend model).

Tests

Host: test_power_final_27 (stock panel table byte-exact against the decoded image, reset-high/fallback order, VGP2 hold balance, DTS/validator contract, switches), test_pmic_sleep (VSYS group reaches stock, mode bit last/first, refuses any selector mismatch, selectors unwritable), test_vproc_sleep/test_pmic_lowpower updated for the admitted VSYS mode bits. Build: kernel warnings unchanged (47, all in the vendor Wi-Fi driver). Platform: validate.py PASS, ELF closure 354 PASS, installed ARM PASS (40 modules), privacy clean, inventory 108 packages, container suite 701 OK (693 + 8).

Owner checks after the flash

Boot, screen on/off several times (panel wakes every time; dmesg | grep -i "panel\|gc9503" has no
stock power-down cmd … failed), and a screen-off longer than a minute.
Play music after the DAC has been idle (first play after a pause; listen for pops), with and without headphones in.
cat /run/y2/power-features.json: panel_stock_suspend, vgp2_idle_off = Y, pmic_sleep = 15;
cat /sys/kernel/debug/regulator/regulator_summary: vgp2-audio use 0 while idle.
One short sleep, then pmic_sleep_state (entries up, vsys_refusals=0, last_changed includes the VSYS bits
0x3c0) and the PMIC capture (0x256 row now matches).
Overnight: start ≥4 h after charging, ≥12 h; compare the late-night voltage slope with 26's (archive 124).