POWER-FINAL-13 — stock suspend power parity (golden master from the Y2's own firmware)

Status: FLASHED, FIRST SLEEP PASSED, USB-AFTER-WAKE DEFECT OPEN (published as Baseline 13, tag baseline-13; see physical result). Sealed as IMPLEMENTATION_READY_PHYSICAL_NOT_RUN. Branch codex/platform-v1-power-final-13 (worktrees ~/Dokumente/Code/worktrees/platform-v1-power-final-13/{Y2Linux,Y2Reborn}), Linux build commit 47c1cf225dc140d151366f9e6309a3a07556fa67 (later test-only tests: commit), Reborn unchanged (71ddb0d). Package out/y2linux-platform-v1-power-final-candidate-13/ (fallback = sealed 12), symbols …-13-symbols/, receipts out/deep-sleep-final/build13-receipts/. BOOTIMG f73c24da3abe5978f2c234bd9501a929415fc156fc5ab79ebc7755390dddc178, Y2ROOT e62cb6c7bab621e6f791359e5b56bba3a21f065d2ba864f4917b3ae808c01915; kernel -power-final-13, root v1.43. Offline: container suite 651 cases PASS (10 host-dependency skips), 32 packaging-host + 27 splash PASS, package/ELF 385/1416/installed-ARM/inventory/privacy PASS. Scope audit: POWER-FINAL-13 entry in docs/planning/roadmap-gap-audit.md (13 branch). Data: docs/validation/data/y2-stock-suspend-golden-20261011.json (13 branch).

Why 13

Defect A on 12 (owner's second sleep 9 s after a Power wake failed in the radio quiesce) left the product
guard pending. Admission matched the guard by build identity only, so every sleep was refused across reboots: archive 98 shows boot 09ce553a awake with the screen dark for 4.2 h, 301 refusals, 78% to 56% (≈4.9 points/h, voltage-derived). The failing command was bluetooth-power off: the reconnect daemon (tools/connectivity/reconnect.c) holds /run/y2/bt-operation.lock while it reconnects after a wake, and bt_control.power() refuses at once.
Drain: 12's first good night was ≈2.2 points/h asleep, against ≈4.9 points/h awake-dark: sleep only halves
the drain. The owner's target is stock parity (≈1% per night, ≈0.1 points/h). No current sensor exists; every number here is voltage-derived and indicative only.

Golden master: the stock FM kernel

Source: evidence-private/20260908-ram-map/fm-decompressed.bin (the Y2's stock kernel, base 0xc0008000) with fm-symbols.json, disassembled with llvm-objdump (armv7). Decoded:

Stock object	Address	Finding
`PMIC_INIT_SETTING_V1` (E2)	0xc04b81f8	80 field writes; 55 differ from the live 12 PMIC (archive 99, read-only debugfs dump). Power-relevant: VPROC sleep voltage set (8 fields), LP_SEL on VMC/VMCH/VEMC3V3, BAT_L/thermal interrupt enables, EFUSE/AUXADC clock PDN. Not power: ADC trim selects (would shift Y2's battery calibration), charger/speaker/start-up fields.
`mt6323_power_gs_suspend_data`	0xc0757cd0	44 PMIC rows; identical to the vendor table the 05 report used.
`mt6582_power_gs_suspend_data`	0xc0759c98	57 SoC rows: clock muxes powered down, INFRA/PERI clock gates, DCM, MAIN/UNIV/MM/MSDC PLL software-enable 0 and AP_PLL_CON3 hardware control, CLK_SCP_CFG 0x3ff/0x11. Never read on Y2 before 13.
`tpd_suspend` / `tpd_resume`	0xc04a41c0 / 0xc04a3088	APT32F wheel: reg6 firmware; 1..19 gets the sleep command reg5=1, else GPIO56 (reset) held low; resume pulses GPIO56 and writes reg5=0. This unit reads reg6 = 0x01. Y2 never slept the controller.
`usb_phy_savecurrent_internal`	0xc04d98f0	Full PHY power-down whenever no cable is present. Y2 only ran the recover half at boot.
`pmic_mt6323_suspend`	0xc04b3dd4	AUXADC hardware mode and VREF18 off in suspend (not adopted: would change Y2's ADC).
`cs43131_early_suspend`	0xc054bed0	DAC GPIOs low and `hwPowerDown` of VGP2 (refcounted). Y2 already keeps the GPIOs low; VGP2 stays on (not adopted).
GC9503 `lcm_suspend`	0xc04a24f4	Display off, sleep in, power-register tail and AF=02, reset left high (Y2: off, sleep in, reset held). Not adopted.
`slp_suspend_ops_enter`	0xc0035bd4	Only `spm_go_to_sleep`; the PLL/clock state comes from the static configuration above.

Changes in 13 (each switchable at run time)

Change	Switch	Restored / non-fatal
Radio quiesce waits up to 25 s for the Bluetooth operation lock; the failure reason names the radio. The guard is cleared whenever the kernel was not entered; only a `kernel_entry` record quarantines.	—	host-tested (`test_product_auto_sleep`, 3 new cases)
Stock VPROC sleep voltage 0.85 V: 8 stock fields applied only inside `y2_spm_enter` (one CPU, IRQs off) around the PCM sleep, VOSEL_ON preset to the live selector, live voltage re-verified, every word restored mode-bit first before the normal PCM.	`/sys/module/pwrap/parameters/vproc_sleep`, state in `vproc_sleep_state`	refusal or failure keeps 1.15 V and sleeps anyway
LP_SEL on VMC, VMCH, VEMC3V3 (stock init). Enables stay unwritable.	`/sys/firmware/y2_pmic/lp_sel`	readback-verified
Stock USB PHY save-current in the y2-usb `suspend_noirq` (no cable, no session); byte-exact restore in recover order in `resume_noirq`.	`.../y2-usb/phy_savecurrent` (counters)	mismatches counted and logged
APT32F sleep command / reset pulse on wake (stock sequence); GPIO56 admitted with the latch preset high.	`/sys/bus/i2c/devices/0-0051/sleep` (bus number can swap)	failed sleep never blocks suspend; wake retries 5×
Maintenance wake 600 s to 1800 s (stock uses a dynamic period).	—	PCM watchdog timer + 30 s
Read-only stock golden capture at every suspend entry.	`cat /sys/devices/platform/10006000.power-controller/golden` (root)	reads only always-on/INFRA blocks

Owner test after flashing

Plugged in: cat the four switches above; golden shows captures=0 and the awake now values.
Unplug, screen times out, leave ≥ 35 min (one 30-min maintenance wake), press Power once. Then read
vproc_sleep_state (entries = restores, failures 0), phy_savecurrent (saves = restores, mismatches 0), wheel sleep (sleeps = wakes, failures 0) and that the wheel scrolls, USB reconnect works, y2_infra/state (restores>0 means INFRA really powers down), golden (entry rows that still differ from stock).
Overnight drain against 12's ≈2.2 points/h. If anything misbehaves, switch the suspect item off at run time
(echo 0 > …) and compare the next night.

Not done (next steps decided by the `golden` capture)

PLL hardware control (AP_PLL_CON3 / CLK_SCP_CFG) and the DIS/MMSYS power domain, eMMC supply removal in suspend, VGP2/DAC rail, panel deep-standby table, PMIC low-battery interrupt instead of timer wakes, AUXADC clocking, and the awake-dark drain (Reborn blanks only the backlight; the display pipeline keeps scanning out).

Physical result (2026-10-11)

First check (archive 100-power13-first-check, read-only): boot 604bc0e6, taint 0, pm_async 0;
vproc_sleep = Y, lp_sel applied on all eight rails, phy_savecurrent enabled, wheel sleep enabled (firmware 1, sleep command), SPM SPM_CPU_SHUTDOWN_INFRA_PDN with 1800 s maintenance. Awake golden: PLL hardware-control bits (APMIXED+0x00c = 0x780000) and CLK_SCP_CFG already match stock. MAIN/UNIV/MM/MSDC PLL are software-enabled, VENCPLL is on, and AP_PLL_CON1 and the INFRA/PERI DCM differ (all candidates for 14).
First unplugged sleep (archive 104-power13-first-sleep-coldstart-archive; device clock = host minus 37.9 min):
host 15:17:30 to 18:39:33, 12,104 s suspended, 7 kernel suspends, 6 clean 1800 s maintenance wakes, Power wake, complete_result_0. 76% to 76% (4.036 V to about 4.007 V at the maintenance wakes) in 3.4 h: indicative only, from voltage.
Defect B reproduced on a clean wake: the gadget enumerated 5 s after the wake (udc configured,
detached=0), but the host received 0 bytes for 3 min. Owner power-off (shutdown_execute), offline charging, power-on (boot 8cdde856, taint 0) recovered it. That lost the in-memory counters (vproc_sleep_state, phy_savecurrent, wheel sleep, y2_infra/state, the golden capture at sleep entry), and the boot's syslog was not persisted past start-up. Next: repeat the sleep, and if USB is dead, read the counters another way before any restart.