Dead ends¶

Approaches that were tried and abandoned, with why. Read these before proposing something similar; if circumstances changed, say which.

Each entry: what was tried · what happened · why it stopped · what we do instead · date.

Note Air¶

Tried: run the Note Air 2's newer Onyx firmware on the Note Air.
Happened: the Note Air 2 has a different SoC (Snapdragon 662 vs 636), UFS instead of eMMC, dynamic partitions and file-based encryption.
Stopped because: kernel, bootloader and hardware support files don't carry over. Moving single Onyx apps over worked for a while, but the launcher must be signed with Onyx's platform key, and patching Android to work around that broke the system.
Instead: a Linux distribution on the mainline kernel.
Date: 2026-03-30 (SoC comparison); app port rolled back 2026-06-08.

Tried: unlock the stock full-disk encryption from TWRP: the vendor key services, every TrustZone listener, matching process identities, SELinux contexts.
Happened: TrustZone creates and loads the disk key only during a normal Android boot; from recovery it refuses every request. Failed attempts could also leave TrustZone so that the next Android boot asked for a password, which meant a factory reset.
Stopped because: the refusal is inside TrustZone. A workaround that decrypted during a normal boot and then started TWRP on top worked, but it depends on Android booting: when an Android change broke the boot, neither Android nor recovery could reach the data, and the device was rescued by turning encryption off.
Instead: /data without Android's encryption; GlassineOS encrypts its own data with LUKS2 (see decision 0006).
Date: 2026-03-28 to 2026-04-20; encryption turned off 2026-06-08.

Tried: patch the stock kernel's e-ink driver so the fast A2 waveform reaches the panel.
Happened: the driver converts A2 to a full update in two separate places; patches that bypassed them crashed the kernel.
Stopped because: the driver is closed (no Onyx kernel source).
Instead: our own user-space e-ink engine on the mainline kernel, which plays the waveform file's real A2 mode.
Date: 2026-03-30; engine on mainline 2026-09-27.

Tried: flash and boot images over fastboot, with the stock bootloader and with several modified ones.
Happened: every transfer hangs at "Sending".
Stopped because: the cause is not in the bootloader code we can change; it is suspected below it.
Instead: write partitions with dd from a running system, or over EDL.
Date: 2026-03-17 to 2026-03-27.

Tried: EDL with the OEM firehose loader for this SoC.
Happened: it loads, then hangs reading the partition table and never writes.
Instead: the "peek" firehose loader.
Date: 2026-06-02.

Tried: a recovery entry sequence from a 2021 blog post (hold power, then press it five times).
Happened: it never worked on this device.
Instead: close the cover and power on (GlassineOS bootloader) for fastboot, or reboot recovery.
Date: confirmed 2026-10-04.

Tried: set the "hashtree disabled" flag in vbmeta to stop verity on vendor.
Happened: no effect.
Stopped because: verity for vendor is switched on by the verify flag in the fstab inside the kernel's device tree, not by vbmeta.
Instead: edit that device tree entry, keeping its size unchanged (a repacked device tree didn't boot).
Date: 2026-04-17.

Tried: count failed boots in the bootloader with UEFI non-volatile variables.
Happened: the device boot-looped.
Stopped because: the variables don't survive a watchdog or forced reset.
Instead: the GPT slot attributes (retries, successful, unbootable).
Date: 2026-03-28.

Tried: linux-firmware's WCN3990 Wi-Fi firmware and board data.
Stopped because: they are built for the Snapdragon 845.
Instead: the device's own files from its modem partition, in a per-board directory.
Date: 2026-09-27.

Tried: CPU frequency scaling with the mainline qcom-cpufreq-hw driver.
Stopped because: on this SoC the frequency hardware must be programmed by the OS, which that driver doesn't do.
Instead: a port of the CPR and OSM drivers (open loop, capped for now).
Date: 2026-09-28.

Tried: cluster idle states in PSCI platform-coordinated mode.
Stopped because: the firmware refuses that mode, so the clusters never slept.
Instead: an OS-initiated idle hierarchy with the stock state IDs.
Date: 2026-09-28.

Tried: blend windows on the GPU before computing e-ink frames.
Happened: about twice as slow: reading the result back costs more than blending, which takes under 1 % CPU anyway.
Instead: blend on the CPU; the GPU computes the panel drive frames.
Date: 2026-09-29.