Files
UniXNU/docs/devices/e5.md
2026-11-10 19:35:01 +00:00

232 KiB

Rongyue E5 (e5)

Enceka/UniXNU-e5

Rongyue E5 is a 5G senior bar phone (feature-phone form factor) with a removable battery, a 3.5-inch 320x480 ST7365P RGB565 panel, 2 GiB RAM, eMMC and a Unisoc UMS9621 (Qogirn6Lite) SoC with 2x Cortex-A76 plus 6x Cortex-A55. The record covers the BootROM-FDL-LK-BootShim-mu UEFI chain, XNU hand-off, framebuffer console, matrix keypad input, seed userland, the read-only eMMC block service and the read-only SD card service (native SD host, default-speed multi-block reads and an ext4 walk over the e5-openwrt image).

Source record: porting.toml, 4643951c8de27ee0a8f4291db4acdca7a46186b7.

This page is generated from the device record.

Area Status
Configuration implemented
Firmware and boot device-verified
CPU and interrupts implemented
Memory and DMA implemented
Storage and root implemented
Display device-verified
Input implemented
GPU implemented
Network planned
Audio planned
Power planned
Image assembly implemented
Diagnostics device-verified
usb-console device-verified
emmc-windows-r257 device-verified

Configuration

implemented

Components: ArmDTPlatform, ArmCpu, SoftwareAES, XnuPrelink, RamBlock.

Hardware and dependencies

  • UMS9621 SoC feed, iPadOS 17.7.11 build 21H461 OS profile, XNU 10063.141.1 source pin, device defconfig.

Configuration

  • board.name = e5; board.soc = UMS9621; board.soc_source = ums9621; board.os_profile = ipados-17.7.11-21H461; schema_version = 2.
  • kernel.defconfig = e5_defconfig; patch_series applies arch, soc, os then device series.
  • Device kernel archives add e5_emmc_build.py (NativeEmmc IOKit storage driver), RamBlock with virtio-mmio bindings, e5_derive_audit.py storage ABI audit from the pinned kernelcache IPSW, SoftwareAES bindings and mockfs fixture.

Procedure

  • Run ./unixnu defconfig e5_defconfig --device <repo> from the UniXNU checkout.
  • Run ./unixnu status --json --device <repo> and confirm profiles, source pins and patch order.
  • CI (kernel-builder workflow) builds kernel.macho from the pinned sources on a macOS runner.

Development findings

  • R305 manualPhone1x touch/Home/app-switcher/layout pass:601callbacks/599good,2coalesced,594event/5timer reads,0bad/recovery/foreign;294changes/23single taps/6drags/2dual releases/all-up/final0.426HID nofail/drop,originalclientdelivery,2433displayframes/30FNVcaptures,900s/fullreadback/allthree37/AndroidA pass. Allautomaticinputoff. Settings entersblack,post-settingsHome untested; currentB2D258C1 logscoveronlyearlystartup, so no uniquecause is claimed. See docs/input-se2-r305.md.
  • R305 frontend identity uses authentic iPhoneSE2 D79AP/iPhone12,8 product data from17.7/21H16 Apple DT. ProductIDc459fdb0eef455588dbc6f587c20a5372792bbf8 matches the current21H461 MobileGestalt table(row57,type0x80bb21e1). Artworkphone/subtype569/scale1 and chosendisplay-scale1 select native320x480 UIpoints. E5 board/UMS hardware, kernel/OS ABI,rootUUID/reservations and hardware compatibility suffixes remainfixed. A whitelisted retained-ADT edit proves unrelated property bytes unchanged; phone layout/runtime mustbe verified.
  • R286 integrates gated E5 HID startup and its archive before launchd. The accepted display/storage implementation and original Apple family instructions are preserved; new input runtime qualification is pending.
  • R277 qualifies a 640MiB DarwinBoot ownership request on this board: 144080 free plus 19760 wired pages at bootstrap, full 600-second runtime, no retained corruption/panic, fresh Android return. board.payload.adt.owned_memory_size now defaults to 0x28000000; the typed property is the only ADT change, within the conventional-memory donor.
  • Arrays replace inherited selections in board.toml; the device list restates every required component.
  • CI rebuilds from identical patch content are not byte-identical; kernel behaviour was validated by flash evidence instead of binary hashes.
  • Patches must be generated against the arch+soc+os-patched tree; anchors against pristine Apple XNU fail where UniXNU series modify the same files (for example arm_init.c gains boot_progress helpers). The payload now requests 512 MiB through payload.adt.owned_memory_size; flash #59 qualifies that setting with kernel 262 and all other inputs fixed.
  • board.toml now publishes a 48-byte development boot-manifest-hash and the E5.Bringup.ImageSet.v1 owner domain for the exact verified OS/SystemOS image pair. The pinned common ADT generator reproduces the board-tested sidecar byte for byte with its saved random seed; all pre-existing properties remain unchanged.
  • V7 adds the ordered e5_display_build.py archive step after storage compilation. It appends the concrete display probe to the already force-loaded E5 archive, verifies the exact original family and every compiled virtual binding, and strips only the owned automatic initializer. CI 336 passes compile/link/prelink and its final linked inheritance/late-init audit. R81 confirms the resulting concrete factory/init/release path with all 173 late samples at a=3 v=1ff s=18a0 r=0; this does not qualify display publication or scanout.
  • R109 diagnostic preparation is isolated on fix/display-submit-trace-r109. The original CI362 scanout attribution is corrected in docs/logs/display-content-audit-r108.json; all 86 resolved patches replay with pinned source inputs, recorded in docs/logs/display-submit-r109-patch-check.json. Only kernel diagnostic code and the panel probe change; payload/firmware/Data acceptance remains pending.

Completion criteria

  • All source pins resolve; every selected component has a record; status shows the intended overrides and patch order.

Current limits

  • Hardware-facing patch changes still require device validation. The earlier general binary-layout-fragility diagnosis was withdrawn after fixed-seed control experiments isolated incorrect keypad MMIO addresses.

Next action

  • Keep source and payload inputs pinned per experiment; change one variable and retain matching flash evidence for hardware-facing changes.

Evidence

Firmware and boot

device-verified

Hardware and dependencies

  • Vendor LK bootloader on the B slot, BootShim raw payload stub, mu-based UEFI firmware (Silicium e5Pkg), DarwinBoot EFI loader, payload FAT volume.

Configuration

  • Boot chain: BootROM to FDL to LK to BootShim to UEFI (4 MiB SILICIUM_UEFI.fd) to UEFI Shell (startup.nsh) to E5BootMenu.
  • B slot boots the payload (misc slot record: B priority 15, A Android fallback 14); watchdog 300 s reboots failed boots.
  • BootManagerLib launches the Shell through an E5 protocol (GUID 9A48E983-AB52-4043-828C-5366DCDADB26) resolved by LocateProtocol at TPL_APPLICATION because DXE event queues never drain without timer interrupts.

Procedure

  • Build the firmware and payload with uefi/build-e5.sh, make-e5-payload.py and mk-boot-payload.py in the e5-xnu repository.
  • Flash with tools/e5-uefi-cycle.sh flash (writes boot_b payload at LBA 8 plus the misc slot record).
  • Collect logs with tools/e5-uefi-cycle.sh collect (Android pstore ramoops zones survive the reboot).

Development findings

  • R286 adds input source/ABI audit records. The existing firmware, transport, memory and storage implementation is preserved; the original HID instructions are hash-qualified and the new input runtime trial is separate from historical evidence.
  • UEFI timer interrupts are not delivered (GIC delivery unconfirmed), so the UniXNU menu counts down by polling and UEFI event notifications at or below the current TPL never dispatch.
  • AON registers are clock-gated: the keypad and AON GPIO banks need set-mirror clock writes (0x64901004, 0x64901018) before any register access or the AXI bus wedges.
  • The Shell must be started from PlatformBootManagerAfterConsole via a direct protocol call; a signalling event queues forever without timer ticks.
  • The paired complete FD and BootShim at e5-xnu b6949f4 boot both CI-306 experiments. The second report observes E5-PREV-CONSOLE: E5TW pos=16307 tail 4096, proving the previous kernel header reaches the actual UEFI reader. Both cycles return to fresh Android boot IDs on slot A after full payload readback verification.
  • CI-314 Data rounds 61 and 62 preserve the entire already tested firmware prefix. Both complete 62926860-byte payloads match flash readback and return to Android A under fresh boot IDs. Round 61 recovers the matching-UUID APFS policy panic; round 62 changes only the ADT via e5-xnu 188f904 and retains live d=100/published/ready checkpoints after the early console wraps. Firmware code is unchanged; the paired reader and collector continue to retain usable evidence. See docs/logs/data-volume-round61.json and data-volume-round62.json.
  • Rounds 63-65 all preserve the same complete firmware prefix and CI-315 boot payload. Each 62926860-byte payload matches full flash readback before reboot. The collector requires changed Android boot IDs and recovers the matching kernel panic in 63/64; the R65 retained ring instead records continued late service activity without panic/ENOSPC. Android A recovers after every test. See docs/data-capacity.md and docs/logs/data-capacity-round63.json through round65.json.
  • The v6 display registration experiment uses the retained r79 payload firmware, EFI files, seed, trust cache and lcd ADT. Only kernel.macho is replaced with CI 335; the complete candidate payload is verified against boot_b before reboot. The family audit JSON records a host observation and does not qualify new firmware code.
  • R81 uses CI336 with 32768 more kernel bytes. The payload helper extends only the kernel chain into 16 verified free FAT16 clusters and adjusts its directory size; firmware bytes and every other file allocation/content remain fixed against r80. Both FAT tables agree, independent mtools reads reproduce all nine files, and all 62926860 flash bytes match before the sole reboot. R81 completes the hardware-watchdog cycle and returns to fresh Android A; retained current-v7 samples and the full unchanged Data hash are archived. This qualifies the fragmented kernel FAT read on this boot, not new firmware functionality.
  • R109 diagnostic preparation is isolated on fix/display-submit-trace-r109. The original CI362 scanout attribution is corrected in docs/logs/display-content-audit-r108.json; all 86 resolved patches replay with pinned source inputs, recorded in docs/logs/display-submit-r109-patch-check.json. Only kernel diagnostic code and the panel probe change; payload/firmware/Data acceptance remains pending.
  • R109/CI363 qualifies the submission diagnostic on device. Full payload readback passes; all 101 retained read samples show F=1:557a87bd while 42413 video-console character calls are suppressed. Original selectors 3/4/5 have zero calls. backboardd PID 56 repeatedly calls selector 6 (SwapWait), with at least 11911 calls, zero successes and last status e00002d5 Busy; selector 12 (RequestPowerChange) is called once and returns e00002c7 Unsupported. The native power/readiness contract is the next target, not an inferred empty scene. Fresh Android A returns after 665 s; full 128 MiB Data SHA c1744ab4 and protected OS/SystemOS metadata remain unchanged. No physical panel observation or UI presentation claim.

Completion criteria

  • Firmware image builds, enters UniXNU, publishes the framebuffer and preserves the documented memory reservations.

Current limits

  • GIC timer delivery in UEFI is still broken; countdown and Shell launch work around it with polling and direct protocol calls.

Next action

  • Root-cause the missing UEFI timer interrupt through GIC distributor/CPU-interface state dumps.

Evidence

CPU and interrupts

implemented

Hardware and dependencies

  • 8 CPUs: cluster0 cpu@0 to cpu@500 are Cortex-A55, cluster1 cpu@600 and cpu@700 are Cortex-A76 (capacity-dmips-mhz 420 vs 1024 in the vendor live device tree); PSCI 0.2 enable-method; GICv3.

Configuration

  • PSCI 0.2 CPU_ON/SYSTEM_OFF used by BootShim and DarwinBoot; XNU boots with the UniXNU ArmCpu/ArmDTPlatform components.
  • XNU kernel timers run (clock_gettime and ITIMER_REAL wakeups verified from the seed clock binary); UEFI-periodic timers do not (see firmware area).

Procedure

  • Compare dumps/live-fdt.dts cpu nodes for cluster layout and capacities.
  • Boot XNU and read the E5UL uplink clock test lines (GTOD, ITIMER, WAITING-ALRM).

Development findings

  • cpu@600/cpu@700 capacity 0x400 against 0x1a4 for cpu@0 to cpu@500 confirms the 2x A76 plus 6x A55 Qogirn6Lite layout.
  • XNU schedules and delivers interval timers on the boot CPU; secondary-core bring-up inside XNU has not been measured yet.

Completion criteria

  • Every enabled CPU reaches its entry checkpoint and timer and IPI handling works on each enabled CPU.

Current limits

  • Enabled-core coverage inside XNU and IPI delivery across clusters are not yet measured.

Next action

  • Instrument XNU CPU_START paths per core and record which of the 8 cores come online.

Evidence

Memory and DMA

implemented

Hardware and dependencies

  • 2 GiB DRAM starting at 0x80000000; vendor reserved regions from the live device tree; kernel, ADT, ramdisk and boot-args staging windows.

Configuration

  • Flash #57/#58 use a 256 MiB owned window with kernel physical base 0xC407C000, boot-args at 0xC819C000 and staging top 0xC8234000. The ADT declares 2 GiB DRAM but originally omitted darwinboot-memory-size.
  • payload.adt.owned_memory_size = 0x20000000 (512 MiB) is accepted in flash #59: loader mem=20000000 and owned=0x20000000, followed by the same root and five successful commands. The UEFI High RAM donor is 0xC4000000..0xF0000000, separate from framebuffer, initrd and top-of-RAM logs.
  • PE_init_platform copies boot_args Video fields; Video v_display is forced to 0 so XNU picks the text-mode console and maps the framebuffer itself.

Procedure

  • Read the DarwinBoot staging debug lines (kernel va/pa, args address, video base) from a collected report.
  • Verify the APFS seed volume mounts from the ramdisk (E5SEED volume UUID in board.toml).

Development findings

  • R305 manualPhone1x touch/Home/app-switcher/layout pass:601callbacks/599good,2coalesced,594event/5timer reads,0bad/recovery/foreign;294changes/23single taps/6drags/2dual releases/all-up/final0.426HID nofail/drop,originalclientdelivery,2433displayframes/30FNVcaptures,900s/fullreadback/allthree37/AndroidA pass. Allautomaticinputoff. Settings entersblack,post-settingsHome untested; currentB2D258C1 logscoveronlyearlystartup, so no uniquecause is claimed. See docs/input-se2-r305.md.
  • R305 frontend identity uses authentic iPhoneSE2 D79AP/iPhone12,8 product data from17.7/21H16 Apple DT. ProductIDc459fdb0eef455588dbc6f587c20a5372792bbf8 matches the current21H461 MobileGestalt table(row57,type0x80bb21e1). Artworkphone/subtype569/scale1 and chosendisplay-scale1 select native320x480 UIpoints. E5 board/UMS hardware, kernel/OS ABI,rootUUID/reservations and hardware compatibility suffixes remainfixed. A whitelisted retained-ADT edit proves unrelated property bytes unchanged; phone layout/runtime mustbe verified.
  • R286 selects the E5 input archive and e5-hid boot argument, with a small report/transition workloop. Declared memory ownership, DMA windows, ADT, firmware, trust cache and media layout retain the R285 values; input runtime and physical tests are recorded separately.
  • R277 qualifies a 640MiB DarwinBoot ownership request on this board: 144080 free plus 19760 wired pages at bootstrap, full 600-second runtime, no retained corruption/panic, fresh Android return. board.payload.adt.owned_memory_size now defaults to 0x28000000; the typed property is the only ADT change, within the conventional-memory donor.
  • R277 candidate requests 640MiB through the existing typed darwinboot-memory-size property. The R275 loader allocated at PA 0xc407c000; adding 0x28000000 ends at 0xec07c000, within the firmware conventional-memory donor 0xc4000000..0xf0000000. The allocator checks the actual map at boot; framebuffer, initrd and top-of-RAM logs remain outside the requested range. All other ADT properties are byte-identical, and the default board request remains 512MiB until hardware acceptance. R277 VM page counters will quantify pressure during LS seeding.
  • The packed ramdisk growth from 291040 to 467184 bytes shifted the staging top by 0x2C000 without affecting boot; staging is size-independent within the reserved window.
  • Memory ownership hand-off follows the boot_args ABI of XNU 10063.141.1 (struct sizes 0x480 verified against pexpert/arm64/boot.h).
  • The complete clean sidecar is parsed before/after the 512 MiB request; all properties except /device-tree/chosen/darwinboot-memory-size remain byte-identical. Kernel, seed, trust cache, EFI files and firmware are held fixed for flash #59.
  • The board-tested E5 image-set identity adds only boot-manifest-hash and its development-domain tag under chosen. All memory request, DRAM geometry, reserved ranges and other ADT properties retain their bytes; the 512 MiB request is preserved in both CI-306 experiments and the common generator exactly reproduces the tested sidecar.
  • V7 adds the ordered concrete display archive without changing ADT memory ownership, framebuffer reservations or Video handoff. Its bounded factory probe requires the registered Legacy size to equal 0x18a0 before constructing an object, adds no private state, releases the allocation and checks instance-count balance. It does not map/write scanout memory. CI 336 compiles the exact size/dispatch checks; r81 measures the original parent size as 0x18a0, validates identity/ancestry and owned metadata, and confirms allocation/init/release with balanced instance count. No scanout memory is mapped or written.

Completion criteria

  • All handoff regions are accounted for and each enabled DMA engine accesses its allocated range with the documented cache policy.

Current limits

  • SMMU is not configured; DMA uses coherently-mapped buffers with explicit cache maintenance.
  • #59 verifies the 512 MiB handoff and unchanged seed workload. High-page stress, complete DMA coverage and memory pressure under original userland are not yet measured; physical DRAM size is not an acceptance claim for a 2 GiB XNU window.

Next action

  • Use #59 as the current seed baseline and retain #58 as the 256 MiB fallback. Measure memory pressure and allocator limits when original launchd/dyld and services are introduced.

Evidence

Storage and root

implemented

Components: RamBlock.

Hardware and dependencies

  • Unisoc sprd,sdhci-r11 eMMC host at physical 0x22200000 (warm-inherited from LK: card selected in TRAN state, RCA 1), eMMC 61079552 sectors (29.125 GiB) with a 70-entry GPT, cache partition at LBA 0xd0a800 size 0x20000 sectors (64 MiB; corrected from the earlier 1 GiB claim), full disk mmcblk0.
  • Independent Unisoc sprd,sdhci-r11 SD host (sdio0) at physical 0x22210000, IRQ SPI 203; SD16G SDHC card, 30244864 sectors (~14.4 GiB), RCA 0x5048, Android runs it at SDR104 4-bit 1.8 V.
  • The SD card currently carries the original OS raw GPT image at LBA 0 (6077580800 bytes, SHA256 e5634d04a8ac35f024845f583394e0a35deb26a04cb0c00286a8674e25a9a4f2). Its APFS partition spans LBA 34..11870241 and OS volume UUID 04769D26-2CA0-4825-B443-E2E5F88300F2. The earlier e5-openwrt ext4 fixture and Android references remain archived.
  • Card power and ADI facts (vendor source audit): vmmc/vqmmc are UMP9620 PMIC LDOs LDO_VDDSDCORE/LDO_VDDSDIO reached over the ADI bus at physical 0x64400000 with the slave window at 0x64420000; LDO enable = PD bit clear (0x21ac/0x21b8), 3.0 V = VOL 0xb4 (0x21b0/0x21bc), write-protect unlock 0x23d0 = 0x6e7f; aon fast-hotplug block at 0x64900520/0xba0 with APB-prot direct-access lock sequence when locked.

Configuration

  • NativeEmmc (scripts/emmc/NativeEmmc.cpp) is compiled into the kernel by e5_emmc_build.py against the storage ABI derived from the pinned kernelcache.
  • v2-v4 probes cover card identification, EXT_CSD, SDMA/ADMA MBR/GPT reads and boot_b/misc readback. v5 publishes the whole disk and IOGUIDPartitionScheme enumerates 70 partitions.
  • v7 published a cache window at LBA 0xd0a800 with an incorrect 0x200000-sector length; v8 corrects it to the actual GPT/sysfs length of 0x20000 sectors (64 MiB), bounded by physical card capacity. The 8 MiB test image inside that partition contains its own GPT and an APFS volume E5EMMC; the physical eMMC GPT is unchanged.
  • board.toml and the ADT now select SD OS volume UUID 04769D26-2CA0-4825-B443-E2E5F88300F2. The separate eMMC recovery baseline #59 retains UUID 238A9EDD-6788-4B2E-BA4E-C1F70AE25B64, kernel 262/v8, a 512 MiB memory request and the signed v9 seed; #58 retains its 256 MiB handoff.
  • NativeSd (scripts/emmc/NativeSd.cpp + SdProtocol.h, personality in patches/xnu/e5-emmc-personality.patch) publishes the SD card read-only. Cold path: ap-ahb sdio0_eb gate, adi_eb gate/reset/mux (idempotent), PMIC LDO enable over ADI (3.0 V both rails), fast-hotplug enable with detect-polarity bit21 SET (cd-gpios flags are 0 on this board), host RESET_ALL + DEBOUNCE DLL backup bits, card hw_reset pulse, 400 kHz 1-bit legacy init (double CMD0, CMD8/ACMD41/CMD2/CMD3/CMD9), capacity pinned to 30244864, 4-bit ACMD6, then the sdio0 chain is observed (not forced) and switched to the Android rpll chain for 24.375 MHz default speed with full re-verify.
  • v15 adds CMD18 multi-block reads (readSectors, 16-sector chunks, v4 32-bit block counts, CMD12 stop, self-compared against single-block CRCs before adoption) used by doAsyncReadWrite; measured 128 sectors: single 36207 us (1.77 MiB/s) vs multi 5995 us (10.7 MiB/s, 6.0x).
  • v16 adds the ext4 read-side walk (pure parsers ext4ParseExtents/ext4CountDirents in SdProtocol.h): sb at LBA 2050 -> group descriptor 0 (inode table fs-block 145) -> root inode 2 extent -> root directory fs-block 4241 with 22 dirents and CRC32 70633ae6, all read through the CMD18 path after the speed-mode verify. The image's superblock first_ino/inode_size fields are corrupt (0/11); the real 256-byte inode stride is proven from the on-disk tables and passed explicitly.
  • Acceptance automation: tools/e5-sd-check.py (e5-xnu) validates every board report from the last 'NativeSd: v1 start' marker — cold-path evidence (gate/pmic/hotplug polarity/CMD8 echo/ACMD41/CMD3/CSD/selected), four reference-LBA CRCs, mode verify (legacy or default with rpll chain), ext4 walk lines, timings, and the full eMMC regression set.

Procedure

  • Clone the existing test image, preserve its volume UUID, create /dev and install the signed v9 PID1, ls/cat/clock and numeric aliases. Generate the trust cache from the actual installed binaries; see docs/emmc-seed.md.
  • Back up and compare the previous 8 MiB cache test image before replacing only that existing test region from Android; verify the entire replacement by readback SHA256.
  • Keep kernel/ADT/EFI/configuration fixed, rebuild FAT and rerun mk-boot-payload.py, flash through e5-uefi-cycle.sh and verify the full boot payload by readback before collect.
  • Require the expected eMMC window, E5EMMC UUID/mount, INIT-START-v9, console fd0, TIOCGETA/ICANON, raw command exit statuses and SH READY/READ-BLOCK in the same report.
  • For SD rounds: capture Android references first (tools/e5-sd-reference.py, ext4 walk recon via adb dd), then flash the kernel with the new NativeSd version and gate the round on tools/e5-sd-check.py PASS over the collected report (see docs/logs/sd-round-log.md for the 17-round root-cause chain).

Development findings

  • R286 integrates gated E5 HID startup and its archive before launchd. The accepted display/storage implementation and original Apple family instructions are preserved; new input runtime qualification is pending.
  • R278 candidate: Data writes use 1..8-sector SDMA CMD25 after an on-boot qualification on the existing spare page at LBA 58728704, outside Data. The spare page and both adjacent guard pages are read twice; a distinct per-word pattern is written/read back, both guards are checked, and all three pages are restored through CMD24 and read back. Failed qualification retains CMD24; failed restoration refuses window publication. Every enabled Data write retains staged bytes in the non-DMA shadow and succeeds only after card-ready, CMD18 readback and byte-for-byte agreement, with ioLock held throughout. A full pre-R278 4GiB Data backup is retained by the host before flashing. Host compilation/link/prelink and mtools payload readback pass (kernel abcfc2d2); device acceptance pending.
  • R277 adds a lock-free VM page census next to E5IO/E5CS4K. The live kernel manages 512MiB (131072 pages), not the full physical 2GiB; measure pressure while the built-in application database seeds before attributing all small I/O to a stalled graph.
  • R275 live observation: the pre-launchd sweep executes in PID 1; pinned ubc_subr.c identifies cs_native4k_prepare as the eager 16K-authentication/4K-digest derivation path. Sampled 1MiB reads take about 74ms with only 1-3us waiting for ioLock. R276 removes unused diagnostic CRC passes after the trace quota (including the discarded CRC on every non-first bulk chunk), retains canary rereads/comparison and locked delivery, and publishes existing native4k read/hash timing counters. Device acceptance pending.
  • R274 holds ioLock across staging, DMA/verification and client delivery on both eMMC window paths and Data I/O; NativeSd Data I/O follows the same rule. The supplied R274 USB capture has three boots with no CAFEBABE, APFS checksum/corruption, code-signature load-error or panic markers. R275 adds a lock-free cumulative E5IO dump and bounded E5READ timing/context samples (first four large requests and every 128th thereafter); context PID is not necessarily the original requester, counters are submitted requests, and hardware validation remains pending. In R274 boot1, two post-early-boot samples advance by 127370 requests and 1220784 sectors (596.086 MiB), so absence of large-read logs does not establish storage inactivity.
  • Physical GPT edits were restored across reboot in earlier tests; the boot-chain mechanism is not isolated. Partition-content writes survive. Current tests leave the physical GPT untouched.
  • The v6 second-device/bare-container attempts did not produce a root mount, including the publisher-thread variant. v7 publishes the GPT-wrapped image through the original matched service and mounts it; the separate effects of GPT wrapping and publication were not isolated.
  • IOKit reads through doAsyncReadWrite are verified for GPT, APFS metadata and seed executables. This proves the tested read path, not every error/recovery case.
  • Flash #54 mounts E5EMMC with old v8; #55 adds v9, real console fd0, ICANON and four standard commands exiting 0. An extra long-argument cat SIGBUS is reproduced before its first trace in #56. #57 fixes the shared static-tool C stack alignment: all five commands exit 0 and the marker file is read. Its original argument pointer is 0x16fdffde8 (8 modulo 16), preserved while the working SP is aligned. SH READY/READ-BLOCK follow with no devfs, AMFI or child exception failure.
  • The #55 through #59 seed-root images match full post-boot readback. #58 changes only the kernel and validates error-first reads, truthful write protection and the measured 64 MiB cache limit. #59 changes only the ADT memory request and passes the same five-command root workload.
  • SD bring-up root-cause chain (rounds 1-13, docs/logs/sd-round-log.md): Android shutdown residue (PMIC LDOs off, host ap-ahb gate off) plus two decode errors and one silent misconfiguration. ADI reads external-aborted because /soc/aon child unit addresses are OFFSETS against ranges 0x64000000 (controller 0x64400000, slave window 0x64420000), not the raw 0x400000/0x420000 from the node names — the cause of five panic rounds. With power on, every response command failed command-CRC with a zero response register: the aon fast-hotplug detect-polarity bit21 was inverted versus the vendor expectation (E5 cd-gpios flags cell is 0; the 0x23 cell is the pin number), and the fast-hotplug machinery gates the response path while it believes no card is present. Setting bit21 made CMD8 echo immediately; rounds 14-15 confirmed the clean driver.
  • SD throughput ladder (round 16): 6.5 MHz single-block (v14) -> 24.375 MHz single 1.77 MiB/s -> 24.375 MHz 16-sector CMD18 10.7 MiB/s with byte-identical self-check; the rpll sdio0 chain left by Android shutdown is alive at XNU (src 195 MHz observed, no fallback needed).
  • The ext4 walk is fully decoded offline against the card (docs/logs/sd-ext4-reference.json + raw captures): root inode 2 (itable+256, extents flag) -> single extent -> fs-block 4241 -> 22 dirents incl '.' inode 2, init (file, inode 537), OpenWrt markers overlay/rom; driver evidence lines ride the same board reports and are checker-gated.
  • The ADT-only image-set identity correction resolves the actual mount-phase-1 exit 66. Original launchd reaches mount-phase-2, init-with-data-volume, fips, keybag, usermanagerd and tzinit task markers. The next genuine panic is tzinit[17] exiting 30 after no writable Data is found and its system-partition fallback returns EROFS. The original OS/SystemOS SD windows remain read-only.
  • After round 60, a read-only 32 MiB SD Data readback has SHA256 42b81cab9b044819aa69d498b2dcf67dcefdcbb9a90651a04891c30d089b00b8. All five checksummed APSB objects and diskutil on the corresponding read-only local image report UUID 30ADC749-351F-4FFF-998A-20E1B683859F; the driver selected stale UUID B5A22774-89CF-41E5-A62B-5AFA6B2710BC. Source now matches the fixture and repeats publication/lookup checkpoints in late sampled SD reads. Host identity, negative controls, failure cleanup and patch application pass. CI 314 builds and prelinks a0a5f1e291b93eda422eeaa945f2c71f622eaabc. Full-readback round 61 confirms RAM publication, the correct UUID at /dev/disk4s1 and entry into RW mount; original APFS then panics on unencrypted Data (apfs_vfsops.c:2276). Its binary checks the /filesystems/DoNotUseSEP property before that guard; round 62 adds only that ADT property through the e5-xnu payload helper. Round 62 adds only that ADT property and all five late samples report d=100, p=published, q=ready; original APFS advances from xid 6 to at least 246 through real RAM Data writes. Later services then exhaust the 32 MiB volume (209-211 free blocks) and report ENOSPC. The card remains unchanged and Android A recovers. Early tzinit exit lines are overwritten, so its explicit exit-zero checkpoint is not retained. See docs/data-volume-uuid.md.
  • Round 63 preparation expands the same Data UUID to a 128 MiB APFS RAM image (262144 sectors) after round 62 ENOSPC. Host image checksum/UUID validation and window tests pass; all 82 ordered patches apply. A development-only startup hook sets and reads back the mounted Data root uid/gid 0:0 and mode 0755 before claiming d=100. The original OS/SystemOS windows remain read-only; CI and board acceptance are pending.
  • CI 315 and rounds 63-65 validate the 128 MiB volatile Data window. d=100 now includes read-back verification of Data root uid/gid 0:0; t=17:0 directly records the committed normal tzinit exit. Round 63 ENOSPC is caused by a roughly 127 MiB watchdogd core; round 64 redirects the core path and retains the same watchdogd SIGTRAP without ENOSPC. Round 65 adds the root-owned standard launchd override for com.apple.watchdogd only: 248 late samples remain d=100/published/ready/t=17:0, reads reach 251392, and no panic or ENOSPC is captured over the hardware-watchdog window. The entire 128 MiB card fixture remains unchanged after each observed run. This is a development baseline with no SEP and a per-service watchdog policy, not persistent Data or complete UI qualification.
  • Display probe v6 adds sampled m/k/l registration, loader-state and full-return telemetry to NativeSd progress. The SD transfer and Data publication paths are unchanged; CI 335 compiles, links and prelinks v6 successfully. R80 retains 202 complete late samples with d=100/published/ready/t=17:0 and completed requests 221312..247040. The full 128 MiB card fixture hash is unchanged after fresh Android A returns. This headless registration test does not qualify UI or persistent Data.
  • V7 adds persistent v/s/r construction, registered-parent-size and postModLoad telemetry. SD transfers and Data policy are unchanged. R81 retains 173 complete samples at d=100/published/ready/t=17:0 with request counters 225408..247424, while the concrete allocation/init/release checks all pass. Fresh Android A returns; the complete 128 MiB Data fixture hash is unchanged.
  • 2026-11-05: r184-r198 hash mismatches name SystemOS disk8s1 inode 50 (dyld_shared_cache_arm64.01), not writable Data. Android buffered and O_DIRECT reads independently reproduce varying wrong bytes at the exact failing pages while eMMC controls pass. A backed-up rewrite of the unchanged original .01 extent restores all four pages (40/40 direct reads), the full 3,303,014,400-byte SystemOS SHA256 b4908cde, and protected metadata. Cold-boot validation is pending in r199; SD media/controller fault separation and long-term retention remain unqualified. See docs/logs/systemos-repair-20261105.json.
  • NativeSd exposes cumulative submitted-request read accounting as RD=512-byte-blocks:operations on the sampled progress line; compare counter deltas with census timestamps to measure SD scan throughput. The card-backed Data fixture is loaded into a volatile RAM device at each boot.
  • The r239 source adds experimental CMD25 writes and an e5-data-persist boot-argument gate for a 1800-second Data-window flush at LBA 19034112 (262144 sectors). Hardware write correctness, crash consistency and cross-boot database retention are unqualified. Back up and verify the complete destination window before enabling the gate; build and verify a fresh payload from this source before a device trial.
  • The opt-in e5-emmc-log-test=<nonce> probe validates primary GPT header and entry-array CRCs, the known boot_b partition bounds, payload and AVB metadata exclusion, and writes one CRC-tagged CMD24 sector at boot_b+61 MiB. It checks R1/programming completion, flushes an enabled eMMC cache, and reads the sector back before publishing the read-only service. R240 passes the single-sector write and warm-reset readback with cache disabled: CMD24 status 00000003, R1 00000900; all other boot_b bytes and the GPT are unchanged. See docs/logs/emmc-log-probe-r240.json. e5-wdt-seconds accepts 60..3600 seconds for bounded diagnostic boots; the default remains 3600 seconds.
  • The e5-emmc-log=<nonce> journal captures console bytes in a 64 KiB RAM queue using a nonblocking producer lock, then commits CRC-checked 512-byte records from a thread call. Eight independent boot slots occupy boot_b offsets 61..63 MiB; early records are retained while later records rotate inside their own boot slot. Producer contention and queue overruns are counted. Storage failures quarantine the DMA buffer and stop the journal. R241 and R242 qualify periodic console writes and cross-boot slot isolation with 86 and 73 valid records, no reported dropped bytes, CRC failures or stream gaps. See docs/logs/emmc-boot-journal-r242.json. Cold power loss, interrupted media writes and full slot rotation remain unqualified.
  • The r244 source adds the opt-in e5-data-wt Data write-through: the Data block service lands every written range on the card (8-sector CMD25 chunks under the host lock, the DMA copy inside the lock) before completing the I/O and completes kIOReturnIOError when the card write fails, turning the RAM image into a read cache. WR=blocks:operations:failures rides the sampled progress line next to RD=. The r239 pre-lock memcpy in flushDataWindow moves inside the lock. r233/r234 established that each watchdog kill leaves the LS recovery file in place so every boot restarts the full bundle rebuild; write-through is the persistence half of letting a completed database survive to the fast path. Card-write correctness, crash consistency and cross-boot retention are unqualified until a device round passes: back up the destination window first, then compare RD/WR deltas with the LS census.
  • R243 (kernel 461e2f06, watchdog 3600 s, journal nonce 20261106, 511 records over 40.2-3635.4 s, zero drops) answers the cold-rebuild question: the LS seeding group never returned to zero (E5LSD raw=fffffffb constant at the single census, t=1483.6 s), SpringBoard never spawned, and the SD host served its last read at t=68.4 s (call 512 of the throttled rw log, ~1.2 MiB total) - the userland runs from pre-staged RAM, so the rebuild is not I/O-bound and no APFS/spaceman ENOSPC appears through the window. At the census one lsd thread (tid 2b5) is still inside the bundle enumeration (CoreServices+0x9515c in the LSDatabaseBuilder region over InstalledContentLibrary +0x53948/+0x499c8 frames) while the other workers sit parked at CoreServices+0x145050: a single long-running scan that does not terminate in one clean hour on this board, versus the same userland reaching SpringBoard in 100 s-5 min under QEMU. The CA-state-gated census fired only once, so E5LSDH/E5THS histories are empty and slow-versus-looping stays open; the next round needs a forced periodic census and the WR= counters to measure Data write volume. See e5-xnu docs/ls-coldrebuild-r243.md.
  • The r245 source (feat/data-media-window-r245) grows the Data window to 4 GiB (8388608 sectors, same base 19034112, ~1.3 GiB card spare) and serves it media-direct: reads stream 16-sector CMD18 chunks and writes land as 16-sector CMD25 chunks under the host lock before their block I/O completes, both counted in RD=/WR=; a failed chunk answers kIOReturnIOError. The 128 MiB RAM image, its boot-time whole-window load/verify and the r239/r244 flush/write-through machinery are removed - there is nothing volatile left to mirror. prepareDataDevice keeps the NXSB gate and cross-checks the container block count against the window (a stale small fixture logs MISMATCH instead of surfacing as ENOSPC later). The card container must be grown to 4 GiB macOS-side first (apfs resizeContainer preserves the volume UUID); the 2026-11-06 pristine fixture backup is e5-xnu uefi/build/data-wt-r244/backup/data-window-128m-r243.gz. E5CAProbe census no longer requires a backboardd task (r243 sampled once in 3600 s) and a new e5_ca_probe_periodic runs it every 60 s from the display paint tick, dumping 2 s later - the series answers slow-versus-looping for the LS scan. Watchdog rounds stay at 600 s. Card-write correctness, crash consistency and cross-boot retention remain unqualified until the device round passes.
  • R245 device round (kernel bf1fe73d, payload 9d72e318, watchdog 600 s, journal nonce 2026110602): the 4 GiB media-direct window publishes cleanly (container 1048576 blocks match, no publication failure) and the boot reaches the SystemOS container about 18 s earlier than r243 (no whole-window load), but the console freezes mid-printf at the SystemOS disk dev_init (~39 s) - the retained ring ends on a half-printed line - and nothing at all prints for the remaining ~560 s until the PMIC watchdog recovers. The journal slot for this nonce holds zero records because the journal starts at 39.5 s, right at the freeze; the ramoops dmesg zones carry the console. No Data-window op is on the timeline (the startup script mounts Data after SystemOS), the swap-file failure preceding the freeze also occurs in r243 which continued past it, and the census would not fire until 90 s. r246 (fix/data-window-instrument-r246) instruments the first sixteen media-direct ops, drops both dataMediaIO directions to the 8-sector command shape, and adds a ~5 s paint-tick heartbeat so the next capture separates console death from system death.
  • R246 device round (kernel 9dc8adef): the media-direct Data reads work - sixteen-plus dmio rd ops (metadata then 8-sector walks to block 40128, zero failures) and the boot mounts the OS volume - but the kernel panics in the SystemOS-container phase (~40 s) in every one of three boots (generations 4-6), each journal slot again holding zero records because the panic lands before the first 5 s journal commit while the panic dump itself only reaches the RAM tee (the E5-PREV-CONSOLE firmware salvage caught the banner tail). r247 (fix/data-window-instrument-r246 + panic-flush) adds e5_journal_flush_panic: the tee panic hook drains the queue to eMMC on the panic CPU with try-locked host access (a stopped CPU may own ioLock; the flush gives up rather than spins), and the journal cadence drops to 2 s.
  • R247c device round (kernel a9fc3670, payload 9da969ac, DoNotUseSEP ADT restored): the deepest 4 GiB boot yet - the Data volume mounts read-write (disk8s1 mount-complete, E5SystemOS data mount result 0), launchd starts, 40 paint-tick heartbeats run to ~450 s and the journal captures 271 records. The round also proves the panic-capture stack end to end: the r247b panic text (E5PANIC: unencrypted data volume is not allowed @apfs_vfsops.c:2276) landed in both the journal and the firmware salvage - root cause was every payload since r244 using the bare CI ADT without the set-e5-scratch-data.py DoNotUseSEP injection (the r245/r246 freezes were the same panic dying mid-print). The remaining failure is the write path: the first media-direct Data writes fail at the command level (NativeSd: mwrite fail lba 19034128 n 8 st 00018001 r1 00000c00 - the r239 multi-block CMD25 shape, never hardware-validated), and the reset-poisoned host later serves EIO to OS-volume reads (fsck binary page-in at ~450 s fails errno 5, launchd emergency-exits on the boot task). r248 replaces writeSectors with per-sector CMD24, the shape the r240 probe proved on this host.
  • R248/r248b/r249 device rounds (kernels 41006f8b/6e3fd277): per-sector CMD24 writes work - zero NativeSd write failures - and the first APFS transactions ever land on the card and survive reboot (six container blocks at 0x1000-0x6000 differ from the pristine grown seed when read back from Android). The boot now mounts the 4 GiB Data volume, starts launchd and then dies deterministically at ~84 s in three consecutive boots (journal commits stop at 84.x s with identical record counts, no daemon ever spawns, a panic banner reaches the firmware salvage). All three panic-capture channels miss the reason: the journal dies with the system, the r247 panic flush produces no output (it runs before the debugger stops the other CPUs and can deadlock on the tee ring lock), and the r249 lock-free bare sector needs a path that never executes. r250 (9a28e7d) adds a fourth channel that cannot fail: the panic hook writes the PANI magic, the panic caller PC and the first 32 reason bytes to physical 0xC6000000/0xC6002000 - plain DRAM that the firmware ring-resume prints unconditionally on every boot with no locks, console or eMMC in the path.
  • R250 device round (kernel 58f27bb8): the scratch channel arms cleanly (NativeEmmc: panic scratch mapped) and the boot repeats the deterministic death one record later (85 records, 86.6 s) - but the firmware reads 0xFFFFFFFF from both 0xC6000000 and 0xC6002000 on the next boot, so not even the initialization zeros survive the reset path. Either the ~86 s death is a hard freeze that never calls panic(), or the region is scrubbed across the PMIC watchdog reset. Next probe for the channel: a rolling heartbeat word written every second distinguishes a silent region (still readable next boot) from a cleared one.
  • R251 (device, pending flash): the r250 0xFFFFFFFF readback is attributed to the mapping class, not the region: the prz console tee writes through ml_io_map() device windows and demonstrably survives the watchdog reset, so r251 maps the panic scratch 0xC6000000+0x3000 the same way (ml_io_map, uncached device), elogs the PREVIOUS boot's scratch words before touching anything (scratch pre - journal-persisted answer even if the firmware read stays dead), arms word0=E5X1 0x31583545 with word1=journal nonce and verifies readback, adds a heartbeat: the 2 s journal pass lock-free stores pass count (word2) + uptime ms (word3) to the scratch before any media work; the panic hook keeps PANI 0x494E4150 + pc + FNV(reason). Interpretation matrix: E5X1+pass count in the firmware line = channel works, death is a hard freeze; scratch-pre readable in journal while firmware reads junk = loss in the LK/DDR-init window; PANI = real panic captured. [Entry reconstructed r253: the original append was truncated mid-write when the session ran out of context.]
  • R251 outcome + R252 + R253 (device+host, local-build era): the first fully local kernel (this Mac, Xcode 27.1, pinned unixnu tree; ce8e6d3c, LC_UUID 003BACCC-E87F-3A00-902E-85B9E6CAC8F3) boots and journals (nonce 2026110602 generations decode 85/86 records). Scratch channel verdict: arm+readback of E5X1 works in-boot, but the zone does not survive reset (next XNU init pre-reads zeros, firmware reads LK-era junk 00024200/7A31000A/01B07A00/03D87011) - 0xC6000000 is rewritten by early firmware/LK between XNU death and any reader; plain-DRAM scratch closed, prz console zone (0xFFF80000/0xFFFB0000) remains the only cross-reset channel. That zone delivered the full r250 death reason: apfs_announce_hash_mismatch disk6s1 inode 54 (= dyld_shared_cache_arm64.05, 115458048 B, extent LBA 13080016) stream offset 86409216 expected 211fbc3f079f3e60cad3591f0dd6e84847b2e34d8fa551ee50d8dd3e4d687b8b got 326cf18c... then bdfb6494... (same offset, DIFFERENT got bytes per boot) -> apfs_vnop_read retval 94 -> check_signature pid 1 error 94 -> dyld cache not loaded -> launchd Library not loaded /usr/lib/libSystem.B.dylib -> initproc panic (kernel UUID 1D514D4E = r250 CI build). r252 forensics: Android raw pull of the whole SystemCryptex container (mmcblk1 skip=6144 count=3150, base LBA 12582912) vs staged source uefi/build/system-audit/userdata/systemos.img - Nov 5 full readback b4908cde41d23e6bc05e6af51a2c9eeb00cf0f863a6a18b1600c96a3fd70414f was byte-perfect, today exactly 9 real differing 4K blocks all inside dyld subcache extents (container bytes 335101952/335392768/340467712/340926464/340934656/341037056/341123072 in .05, 1813651456 in .33.dylddata, 2040143872 in .35), every differing byte a SINGLE-BIT flip (XORs 0x10/0x40/0x08/0x80/0x02) -> marginal media cells reading unstably under XNU bursts, degraded after the r248+ write-enabled boots; XNU got-hashes exist nowhere on card. Repairs: 9-block rewrite (readback x5 == source) still failed next boot at offset 80969728 (whack-a-mole) -> FULL container rewrite from the staged source (dd seek=6144 conv=notrunc, 194 s) with TWO full readbacks == b4908cde. Boot outcome: boot-8 hit the same wall (check_np failure, journal ends 15.5 s uptime at DPU t=310 ticks; paint tick = 50 ms so E5HB %600 = 30 s, E5DPU %100 = 5 s - counters are TICKS not seconds), but boot-9 RAN PAST THE WALL: launchd alive (E5PROC pid=1 exit=0 active=1), 20 processes seen, 60 services registered (E5LIVE seq=3), journal 210 records, DPU ticks with live panel CRC to t=2810 (140 s), then silent death at ~140 s (no records after; ring clobbered by the Android fallback so panic-vs-freeze unknown). r253 change (6a7b8d0): read verify-retry in the NativeSd base read path - every 8-sector CMD18 and single-sector chunk is re-read and delivered only when two consecutive reads agree (3 reads max, last wins); RV=chunks:unstable rides the rw-done census. Kernel f6ea63cde7b2e321c9035f3606df1566b4343a914b4fd9fde9cc89f8ad41ff1c, payload e5-uefi-r253.bin (e50bc10b9c3d513b, 62926860 B). Pending flash outcome.
  • R253 + R254 (device+host): media ruled out, 300 s watchdog found. r253 (6a7b8d0) added read verify-retry (two-agree delivery, 3 tries) to the NativeSd read loop: zero 'rv unstable' lines in the 208-record gen-10, delivered chunks pairwise-stable, so post-rewrite media reads are clean. r254 (f6a67ab) traced big reads: gen-12 journaled 726 records (511 retained), a streaming sequential scan of 1 MiB reads (rw+ exit full 2048/2048 each, LBAs 14709168->14715312 stepping +2048 = SystemOS container byte ~1.09 GiB, the LS bundle-rebuild scan), last record 'rw+ enter 2048 at 14715312'. All four long boots (r251 boot-9, r253 gen-10/11, r254 gen-12) die at uptime_ns 338.3/338.5/338.8/339.4 s - a deterministic timer, not media, not a wedge. Calibration: NativeEmmc arms the UMP9620 watchdog at driver start (boot-arg e5-wdt-seconds=300, census line 'wdt window 300s'), driver start is ~38 s into kernel boot, 38+300=338 s kill; nothing pets it. The system at death was healthy: reads completing, launchd alive (20 procs), panel painting live CRCs. Tick scale corrected by uptime_ns correlation: 1 paint tick = 120 ms (8.3 Hz), E5HB every 600 ticks = 72 s, E5DPU t= every 100 ticks = 12 s; uptime_ns is authoritative. Fix flashed: payload toml e5-wdt-seconds 300->3600 patched in-FAT (file 789->790 B, same cluster, dir size field patched, payload e5-uefi-r254-wdt3600.bin ec0e88b3, source uefi/Platforms/Unisoc/e5Pkg/payload/unixnu.toml synced). Expect the kill at ~3638 s uptime, giving the rebuild scan a full 3600 s (r243 precedent: lsd scan still active at 1483 s, 25x slower than QEMU). Next wall if the scan finishes: the SpringBoard/lsd CoreServices+0x1029c8 dispatch_group_wait era. The SD card has proven marginal cells (9 single-bit-flip blocks rewritten in the r252 full container restore) - replacement advisable long-term.
  • R255 (device+host, 2026-11-07): first live-watched boot over native USB CDC ACM + first scan-rate measurement. Payload d7a8377f built from the accepted USB artifacts (kernel 185e7376 = r254 source + USB console, firmware d2603981, wdt 600, nonce 2026110701, e5-usb=1), flash readback verified. USB capture connected 0.7 s after start ('E5 XNU log'), 552576 B over 589.2 s dropped=0, disconnect = the watchdog kill; kernel.log is now the authoritative stream (journal slot 2026110701 retained only 511 records as a partial ring). The LS bundle-rebuild scan ran the full window at a steady 3.25-3.37 MiB/s: 2045 completed 1 MiB reads (rw+ exit full), LBA 12633360 -> 16805536 with a read in flight at death. The scan restarted from the SAME LBA 12633360 as r254 and re-read the identical region - deterministic, no cross-boot checkpoint; r255's 600 s covered exactly 2x r254's 300 s. Census every 60 s (E5CA g=2..18) all counters 0 (no SB, no transitions), E5PROC launchd only, panel painting with a single static CRC 11b0af80 the whole window. System healthy at death - the watchdog is the only killer. Projection: SystemOS container spans ~LBA 12414384..18865584 (3.07 GiB); remaining ~1.0 GiB needs ~310 s more reading, scan ends ~kernel+950-975 s, then lsd DB write + SB/backboardd chain (unmeasured on board; QEMU ~100 s post-ready, board historically 10-25x slower). Watchdog policy hardened this round: payload default 300 s, ceiling 600 s, longer windows require the user's explicit E5_WDT_AUTHORIZED=<seconds> (e5-xnu 5a10f17); tools/e5-scan-project.py (e5-xnu) parses rw+/E5CA/E5DPU from a live capture for rate+projection. r256 needs a >600 s window (recommend 1800) to cover scan end + first SB chain attempt, pending user authorization. See e5-xnu docs/ls-scan-rate-r255.md.
  • R256 (device+host, 2026-11-07): first SpringBoard spawn on device, then watchdogd-crash panic ended the boot. Payload aa55cb47 (user-authorized 1200 s window, nonce 2026110702) flashed with both USB products captured live; the scan restarted deterministically at LBA 12633360 at 3.2-3.28 MiB/s exactly as r255. After the scan window: runningboardd pid=26, SpringBoard pid=27 spawned (E5PROC seq=23), lsd registered the SB bundle (E5LSDI pid=27 t=/System/Library/CoreServices/SpringBoard.app) - the first on-device SB spawn. APFS then logged obj_checksum_verify/obj_read failures whose cksum/oid bytes decode to ASCII console-log text (deterministic wrong content delivered by SD reads, mechanism unresolved; 9 rv-unstable LBAs). watchdogd crashed 3x (E5EXC exc 6, faulting struct read as all zeros) and XNU panicked: watchdogd[103] exited -- exit reason namespace 2 subcode 0x5 (panic at ~kernel+970-1000 s, before the 1238 s hardware kill; USB capture the only record). Root cause of the panic: kern_exit.c proc_handle_critical_exit panics when a POSIX_SPAWN_PANIC_ON_CRASH process dies; boot-arg panic_on_crash_disabled=1 disables it for non-initproc (TUNABLE exists in xnu-10063.141.1, no kernel patch needed). E5CA census all-zero whole run, panel CRC static - no UI atoms yet. Artifacts: e5-xnu uefi/build/sb-r256/.
  • R257 (device+host, 2026-11-07): eMMC window migration round; device bricked at first cold boot after GPT shrink edit, recovered via fastboot. Online f2fs shrink of userdata p70 (5729280 -> 3107840 4KiB blocks) via openwrt-tools f2fs-online worker; GPT end-LBA edit on both copies; SystemOS+Data dd-ed to raw LBAs 40103936/46555136; all readback-verified while Android stayed up. r257 payload (kernel 9889ec6e, kext eMMC windows @41edabe) flashed to boot_b, slot B armed; on reboot the device never enumerated any USB product and stayed dark; LK fastboot later showed a broken partition view (read miscdata fail / get partition info fail / boot ctrl unreadable). Prime suspect: this unit-s LK rejects the edited GPT at cold boot (U30 Pro and changxing60 units cold-booted the same edit); boot_b likely never loaded; kernel dark-panic not excluded. Recovery: fastboot flash partition + flash misc from pre-shrink backups -> Android A up in ~90 s; backup GPT restored via adb dd; vold resize.f2fs regrew f2fs 12 -> 22 GiB; /data and apps intact; both raw window images re-verified byte-identical after the regrow. Policy: no more GPT edits on this unit; windows live inside p70 free space and must be hash-verified before every XNU round; r258 = same payload under the original GPT to disambiguate GPT-vs-kernel. Artifacts: e5-xnu docs/emmc-shrink-incident-r257.md; uefi/build/emmc-shrink-r257/backup/.
  • R258 (device+host, 2026-11-07): first eMMC-windows boot; kernel exonerated, GPT confirmed as r257 dark-boot cause. Payload be1388de (kernel 9889ec6e, panic_on_crash_disabled=1, wdt 600, nonce 2026110704): both USB products enumerated (UEFI +24.9s, XNU +45.2s), NativeEmmc liftClock succeeded (div 8, 4-bit, 16-sector 3525us ~2.3MB/s effective; caps base read 0 so ladder fell to [8,4,2] and the KB/s print is a base-0 display bug), SystemOS window published as disk3 (806400 blocks), Data window as disk7 (1048576 blocks), SD card self-refused and armed the watchdog, boot_b payload ran. New failure: disk7 APFS mount saw journal records as checkpoint objects - obj cksum 0x0031474f4c423545 LE-decodes to the E5BLOG1 record magic, oid high word = the boot nonce 2026110704, xid = journal gen 25, object types 0x11/0x21/0x31 = record seq 17/33/49 = first records of flush passes 2/3/4; e5_systemos_device 30s wait expired (SystemOS disk3 dev_init but no slice/mount in time) -> panic E5SystemOS: preparation failed: 2 @kern_exec.c:8341 (x2, wdt killed at ~638s, B tries exhausted, Android self-recovered). Android-side forensics: Data window media content byte-identical to the staged image (container block 8 = valid NXSB copy), journal area slot reads return ARM code (stale boot_b payload bytes) - media is fine, the driver served wrong bytes. Reads that hit during journal flush windows returned the just-staged journal records; every readMulti reported success (zero poison-tail failures, zero rv-unstable - the two-agree verify saw the same wrong bytes twice, deterministic per LBA). All paths hold ioLock on inspection, so the aliasing mechanism stayed hidden; r259 diagnostic/mitigation: window I/O moved to the second half of the shared 8KiB DMA page (journal staging keeps the first 4KiB), per-call LBA+crc32 trace for the first 64 window reads, SDHCI ARGUMENT/TRANSFER_MODE readback drift check after every CMD18, probes moved under ioLock, one E5JRN pass marker per flush. Artifacts: e5-xnu uefi/build/sb-r258/ (capture 29984B kernel.log, journal boot-25 slot), expected-crcs table uefi/build/sb-r259/expected-crcs-data.txt.
  • R259/R260 (device+host, 2026-11-07): r259 diagnostics kernel + r260 journal-off self-test both data-aborted in liftClock before any window work - r260 reproduced the identical signature (abort right after the caps print, far=regs+0x3000, x22=0xcafebabe; pc shifted 0x44 by the added self-test code). Root cause: the r259 buffer split moved window I/O to the second 4 KiB half of the 8 KiB DMA page but liftClock still probed with 16-sector readMulti - CMD18 SDMA wrote 4 KiB past the page end and corrupted adjacent kernel memory (r258 never hit this: its window buffer was the full page). Fix 858522d: readMulti hard-clamped to n<=8 (the half-page ceiling every window path already uses), probe at 8 sectors. Lesson: SDMA length must be checked against the ACTUAL buffer span, not the nominal page. Artifacts: uefi/build/sb-r259/r259/capture/kernel.log, uefi/build/sb-r260/r260/capture/kernel.log.
  • R261 (device+host, 2026-11-07): fix verified + journal-off verdict: corruption persists, media exonerated at the fault LBA. Payload 9563f060 (kernel 471eaca6, journal dropped via --drop-arg e5-emmc-log, panic_on_crash_disabled=1, wdt 600): liftClock succeeded (div 8, 8-sector 1790us ~2.3MB/s), all three eMMC windows published, SELFTEST t0/t1/t2 CLEAN bad=0 over the first 1 MiB of os/dm/spare, APFS mounted BOTH windows with zero corruption complaints - then boot went further than ever: E5SystemOS union mount succeeded (disk8s1 SystemCryptex volume, uuid 8846FE78, dyld cache header verified, dyld_flags 0x8) and Data mounted rw (disk6s1). launchd exec died: apfs_announce_hash_mismatch inode 49 (the dyld_shared_cache_arm64 480KiB stub) file offset 475136 len 16384 - expected sha256 bc402e98... got ea20f5bb... -> dyld cache registration failed -> Library not loaded libSystem.B.dylib -> initproc failed panic (panic_on_crash_disabled does not cover initproc). Forensics: inode 49 tail 16 KiB located UNIQUELY in systemos.img at byte 25272320 = eMMC LBA 40153296 (window chunk 6170, ~24.1 MiB in - beyond self-test coverage); Android dd of that exact LBA returns the CORRECT bc402e98 content; the wrong ea20f5bb content exists NOWHERE in systemos.img, boot_b-64MiB, or Data-window-head-64MiB (aligned+sliding 4K/16K sha256 prefix scans) - the wrong bytes are not stale media content, they are a composite or DRAM-sourced mix. zero rv-unstable, zero xfer drift, zero poison failures all round: the wrong bytes were served consistently to both verify reads. r262 = corruption map (sparse 1-per-1MiB CRCs across both windows + dense fault region + wrong-chunk 32B hex dump). Artifacts: uefi/build/sb-r261/r261/capture/kernel.log, bootb-64m.bin, data-head-64m.bin.
  • R264 (device+host, 2026-11-07): kernel 37c70bad payload a5decb3c wdt 300 journal off. SD media publication gated off (e5-sd-publish). FIRST BOOT WITH ROOT ON EMMC: Got boot device E5OsRootDeviceEmmc, BSD root disk4s1. Data window published and mounted rw (disk6s1) after head restore. Full prep chain green: union mount disk8s1, dyld cache header verified, dyld_flags 0x8. NEW FAILURE: panic Process 1 exec of /sbin/launchd failed errno 8 (ENOEXEC) kern_exec.c:8444. Host image /sbin/launchd is a valid arm64 MH_MAGIC_64 (compressed on media, signature scan misses). Cryptex has no /sbin so union cannot shadow launchd. Zero rv/drift/FAIL lines. SELFTEST t1 CLEAN bad=0 (dmh+osf+osd+sp + oss 1536 chunks before deadline bail). Roaming-fault pattern: corrupt-read sites moved from cryptex (r258/r261/r263) to root volume launchd exec. Media re-verified from Android chunked: all windows intact. r265: three-agree reads with shadow delivery, 4096-call wr trace, eMMC cache-off CMD6 byte 33.
  • R265/R266 (device+host, 2026-11-07): r265 v50b 4dc13434 three-agree reads: root eMMC ok, SELFTEST CLEAN, zero rv3 disagreements, but prep failed EARLIER with ENOENT at kern_exec.c:8341 - 3x read cost made cryptex APFS mount miss the 30s IOMedia helper deadline (late wr #227-228 os 29db080 show cryptex still mounting). r266 v51 d22315df two-agree + shadow-copy delivery restored full chain: root eMMC, union+dyld+data mounted, then load_init_program -> /sbin/launchd -> ChunkCompression.cpp:618 decode failed for compression algorithm 0x205 -> decmpfs_fetch_uncompressed_data err 22 -> errno 8 ENOEXEC -> initproc panic (identical in r264). STORAGE EXONERATED COMPLETELY: all 8 launchd compressed chunks (wr #452-459, LBA 0x28bf470.. OsRoot+5.1GiB) logged CRCs match osroot-bare.img byte-exactly; whole-log sweep: all 187 OsRoot reads image-perfect. Cryptex reads showed logged-crc != image at 86/137 sites with content = image bytes from exactly 0xF080 sectors lower (6 independent hits) - now understood as f2fs residue, see R267.
  • R267 (device+host, 2026-11-07): TWO root causes closed. (1) NativeEmmc.cpp windowRead multi-chunk delivery bug: data->writeBytes(c*512,...) wrote each 4KiB chunk at descriptor byte c*512 instead of c*4096 - chunks overlapped, so any count>8 read delivered 1/8-sampled garbage while every chunk CRC (logged from winD8 pre-delivery) stayed perfect; single-chunk reads (self-tests, superblocks, launchd chunks) were correct, which is why self-tests passed and APFS b-tree bulk reads corrupted decmpfs metadata into bogus algorithm 0x205. Fix 908144d, kernel v52 61143108824312dd6572aeac56eb77f7fcb86f2a7a98a3b21b35cd1df18444ec. (2) Windows live INSIDE p70 tail which Android vold resize.f2fs regrows to partition end at EVERY Android return (no fstab flag; Android 13 behavior): post-r266 chunk scans show osroot and systemos windows MASS-rewritten on media (91/92 and 50/50 64MiB chunks differ) while their 64KiB heads stayed intact - resize+GC rewrites gigabytes across the freed tail; r256-r266 roaming faults (Data-head zeros, spare zeroing, -0xF080 shifted bytes = stale chunk-file remnants) all trace to this. Defense per round: full re-place -> f2fs shrink to 8GiB -> flash B_TRIES=7 (LK allows 2..7) so the device stays in XNU boots with no Android return between attempts; preflight window verify gates every flash.
  • R267 (device+host, 2026-11-07): LAUNCHD RUNS. Kernel v52 61143108 (writeBytes c*512->c*4096 fix 908144d), payload 0ea441fe trial 2026110713 wdt 300 journal dropped panic_on_crash_disabled=1, B_TRIES=7 session (3 boots captured). Operational: f2fs shrink itself zeroes the freed tail synchronously (eMMC applies F2FS_IOC_RESIZE_FS discards immediately) - correct order is SHRINK FIRST THEN PLACE; all four windows re-placed chunk-verified (osroot d8808bea, systemos b4908cde, data d8b88ba4, spare a88e884d) post-shrink. EVERY BOOT: full chain green - OsRoot root, union mount, dyld header verified, Data mounted rw, launchd exec SUCCEEDED (reads continue past the 8 compressed chunks #452-459 with identical CRCs, then dyld/lib content reads #460+), shared_region check_np warnings (launchd alive), LS cold scan STARTED on the cryptex (rw+ enter 896 at 43937232 = +23.7MiB, same offset family as r256), scan reached ~+55MiB, ST t1 CLEAN bad=0, wdt 300 clean kill, NO PANIC ANYWHERE. Non-fatal anomaly: nx_checkpoint_find_highest_xid disk8 'object type 0xcafebabe @ checkpoint descriptor block 30' on first cryptex contact every boot - recovers, mounts, runs. Scan rate ~0.21 MiB/s (31.6MiB in ~150s) -> 3GiB cold scan ~4h on eMMC: unreachable in any wdt window. NEXT: pre-seed LS database via QEMU (same IPSW 21H461 cold rebuild completes ~100s there per unixnu.com/devices QEMU virt record), transplant Data volume with DB, skip cold rebuild on board. Artifacts: uefi/build/sb-r267/r267/capture/kernel.log, docs/emmc-launchd-decode-r267.md.
  • R268 (device+host, 2026-11-07): bulk-read path + throughput root cause. Kernel v53 53336e391b002335c5eab99c7c7e8fe76b9acde6d3250cd412dbe29ccd2f2440 (b6ac61e): dedicated 64KiB IOBufferMemoryDescriptor bulk buffer + readMultiBig (n<=128 CMD18 SDMA, poison/drift/CMD12 discipline) + windowRead bulk path (>=64 sectors: 128-sector chunks, first-chunk two-agree CRC canary, per-chunk wr#ob CRC logging) + liftClock rewrite (card CMD6 EXT_CSD 185 HS switch + 4-bit + fastest-first SDCLKFS ladder). Payload becf3c8e trial 2026110714 wdt 300 B_TRIES=7. RESULT per boot (deterministic): HS switch y, fast mode div 2, scan bulk read rw+ 896 at 43937232 COMPLETED (r256/r267 never finished it) then rw+ 2016 in flight at kill; SELFTEST t1 CLEAN; no panic. BUT 128-sector probe = 27.6ms = 2.33 MB/s — IDENTICAL per-byte rate to r267 div8 8-sector (1.79ms/4KiB = 2.25 MB/s): divider writes and HS did nothing. ROOT CAUSE (from 'inherit' log line): LK leaves the host at clk=0x0207 (SDCLKFS=2 + clock-source bits 0x07 = FAST parent) hc1=0x20 (bit5 = 8-BIT bus; DMA select bits[4:3]=00 SDMA) with vendor DLL tuned (dll 0xa0000102 dly 0x3434b760 off 0x00202000) and reads tens of MB/s in that state; our identify re-init writes 0xFFC1 whose low byte muxes the SLOW parent (source bits 0xC0), and setDivider preserved the CURRENT source bits — every rung divided the slow parent. Also learned: caps base field reads 0 (invalid, matches E5BootMenu.c comment 'do not infer Hz from the capability field'). r269 fix (ee2fc3a, kernel v54 68061786b61d4774): snapshot inhClk/inhHc1/inhDll/inhDly/inhOff before identify; liftClock restores DLL regs, CMD6 183=2 8-bit + HC1 bit5 (4-bit fallback), divider rungs on the INHERITED source bits starting at LK's own SDCLKFS. QEMU pre-seed kept as fallback (no longer primary). Placement ops: 256MiB chunks (~19min all four windows), md5sum not md5 on device.
  • R269 (device+host, 2026-11-07): CLOCK FIX LANDED + STORAGE FULLY EXONERATED + LAUNCHD RUNS + 3-MIN SHARED-CACHE MAP FOUND. Kernel v54 68061786b61d47745f55dbb5ffb316f3ab17dd5e31dadbca296e2d86baa612f1 (ee2fc3a): identifyAndRead snapshots LK inherit state (clk 0207/hc1 20/dll a0000102 dly 3434b760 off 00202000) BEFORE any host write; liftClock restores DLL regs + CMD6 183=2 8-bit + HC1 bit5 + SDCLKFS ladder &#123;LK's 2,4,8,16,0xFF&#125; on INHERITED clock-source bits (identify's 0xFFC1 write muxes slow parent - old setDivider kept dividing it; probe now 'fast mode div 2 src 00 8-bit; 128-sector in 48 us'). Payload 7af31622 trial 2026110715 wdt 300 B_TRIES=7, 3 boots. RESULTS: (1) bulk streaming now ~17 MB/s - the 3GiB cryptex scan completes in ~3min per boot (RD=65455 ops/6.4M sectors per boot, 128-sector CMD18 chunks); (2) ALL 657 logged ob-bulk-read CRCs byte-perfect vs images (window-aware verify: osroot-bare.img base 32018432, systemos.img base 43888768) - delivery PROVEN clean; (3) SELFTEST t1 boot1 CLEAN 75ms, boots 2-3 bad=43 - ALL 258 mismatches in dms (Data sparse area) with got32 = APFS file content (asset names, format strings) = EXPECTED live-system writes to the rw Data volume (fsck/mount/lsd DB), NOT corruption; (4) launchd exec + runs + boot tasks progress (mount-phase-1, data-protection, detect-installed-roots, commit-boot-mode); (5) xpcproxy 'ChunkCompression.cpp:404 invalid resource map' x26 per boot - NON-FATAL for most services (tasks ran) but unresolved; (6) THE PRE-LAUNCHD 3-MIN SCAN: between 'load_init_program: attempting /sbin/launchd' and launchd hello, the ENTIRE cryptex container is linearly read (+23.7MiB to +2.8GiB, same start LBA family as r256's lsd scan) = dyld shared-cache mapping reading the cryptex cache file (decmpfs chunks); with wdt 300 launchd lands at +3:57 leaving <60s for services -> SB unreachable in window. r270 (running): SAME kernel, wdt 600 (standing cap; 300 was m02125 verification-phase only), B_TRIES=7, NO re-placement (all heads verified intact; Data volume dirty state intentionally persisted so lsd DB progress accumulates across B-slot boots - the persistent-Data bootstrap design). Artifacts: uefi/build/sb-r269/r269/capture/kernel.log (1.25MB, 3 boots).
  • R270 (device+host, 2026-11-07): SERVICE CRASH LOOP = WRITE BUG. wdt 600 B_TRIES=7 no-replace round (payload 30b6a16a, kernel v54 68061786): root-on-eMMC + Data rw mount green, launchd +3:57, but every service (biomed, ManagedSettingsAgent, contextstored, audioaccessoryd, ReportCrash) crashed paging dyld-GENERATED dyld_shared_cache_arm64.52.dyldlinkedit on disk6s1 (ChunkCompression.cpp:564 resource fork offsets out of order -> decmpfs err 22 -> E5EXC corpse loop). ROOT CAUSE: NativeEmmc writeSingle wrote always winD8[0:512) while dataMediaIO staged n<=8 sectors -> every multi-sector Data write replicated sector-0 content across all LBAs (bug since r257 port). FIX commit 6153668: writeSingle(lba, srcOff) with winPhys+srcOff. r271 = fix + clean Data re-place (d8b88ba4 seed, 16 chunks verified). DISCARD-STORM discovery: f2fs shrink discards 14GiB tail lazily; cheap eMMC erases rewrite-later blocks for ~20+ min after shrink (osroot chunk22 verified OK then re-zeroed 6 min later) -> canary + 3-clean-scan gate before placement; shrink-then-place order; minimal Android dwell between rounds.
  • R342 SpringBoard/Metal prerequisite fixes a duplicate-UUID Data selection bug: R341 mounted the 8182-block RAM seed although NativeEmmc also published the 1048576-block eMMC container. e5_data_device now matches UUID plus the E5DataDeviceEmmc ancestor and rechecks 4GiB geometry. R342 logs SELECT PASS for /dev/disk8s1, APFS block count1048576 and mount result0; all GPU offscreen controls pass. Full closeout is complete-data-verified: runner rc0, all three 37/37 window checks, Android A recovery and locally SHA256-verified 4GiB Data-after 239ddc6fdb6d21c5e99723e5e23c56fbbd96de4ba744b182ceff0f1c2da8351c. Persistent Data bytes changed from the pre-run backup. UI presentation remains unaccepted; see docs/logs/gpu-r342-evidence.json.

Completion criteria

  • Host: SDHCI register, clock, card mode and DMA facts match the board inventory for both hosts.
  • Reads: eMMC capacity, GPT and known blocks are correct; all exported windows stay within their containing partition. SD capacity, CSD, RCA and four reference blocks match the Android capture; the ext4 sb/gdt/inode/dirent chain matches the golden walk.
  • Root: the configured volume mounts and launchd starts. Persistence: writes survive reboot within the allocated Data region.

Current limits

  • XNU eMMC/SD writes remain unsupported. The eMMC recovery volume is a custom APFS seed. The original SD OS image mounts and reaches launchd/libignition with full checker PASS in round 27; Preboot entitlement handling, SystemOS/dyld-cache attachment and writable Data remain unqualified.
  • v8 passes thirteen host completion/error and six window-boundary cases; #58 mounts the same root and runs all five commands with the corrected 64 MiB read-only window. Hardware CRC/timeout injection and failed-request recovery remain untested.
  • Five batch commands and tty setup passed in #57/#58; physical keypad and serial command entry remain separate acceptance items.
  • NativeSd is read-only at 24.375 MHz default speed (SDR50/SDR104 + DLL tuning not attempted; Android reference is 208 MHz SDR104). CMD18 chunks are capped at 16 sectors within one 8 KiB DMA page; multi-block is adopted only after CRC self-check. The ext4 walk covers depth-0 extent trees and one directory block per extent — enough for this image's root; deep trees/hard links/xattrs are out of scope. The card is treated as a fixed golden fixture; a rewritten card changes the checker expectations.

Next action

  • Complete r199 with the unchanged r198 payload and Data after SystemOS repair, then recheck known failing pages after the reboot before attributing display changes to a kernel fix.
  • Preserve the paired e5-xnu baseline-data-r65 boot and 128 MiB Data fixtures, with the no-SEP property, core-output control and explicit watchdogd development override. Data RW/root ownership and tzinit exit zero are qualified; do not regress to a mismatched old 32 MiB kernel/image pair.
  • Qualify the next actual userland or hardware dependency: late R65 logs show missing AppleKeyStore connectivity, missing Bluetooth DT/address and noncritical liveactivitiesd crashes. Full UI and service health remain open.
  • Implement native IOWatchdog/power and persistent Data separately; the current physical-card driver stays read-only while Data writes go to RAM. Keep Android A and exact destination backups available.

Evidence

Display

device-verified

Components: GopFramebuffer, AppleKeyStoreState.

Hardware and dependencies

  • ST7365P 320x480 RGB565 panel driven from LK; UEFI GOP from the inherited framebuffer; XNU boot_args Video hand-off.

Configuration

  • The original Legacy framebuffer family retains source fences and ownership while converting full-screen BGRA or validated linear b3a8 to the firmware-owned320x480 RGB565 buffer.
  • Use backboardd CAEnableDeepFramebuffer=false for this SDR panel. R379 confirms native BGRA desktop presentation; R385 confirms actual Settings content.
  • The presenter advertises Mach FrameInfo transport0 with count1 and boolslot336=false; r388-r390 prove this advertisement (not the reports, which remain unimplemented in slot337) is required for the desktop composite. A protected native-list walk delivers only initialized references, clears the live-port mask after unregistration, and retains the existing100ms software-notification timer. No hardware-vblank timestamp is claimed.

Procedure

  • Boot the payload, wait for the seed shell, and read the E5FB lines from the collected report while watching the panel.
  • Run the E5KINIT read-back from the matrix keypad for the keypad-side sanity values.

Development findings

  • R388-R390 pin the FrameInfo advertisement as load-bearing: opaque213 answering count1/Success (empty64-byte key, transport0, still no completion reports in slot337) is REQUIRED for the desktop composite. R388 with e5-frameinfo=0 (count0/Unsupported) boots identically to R387 then stalls the BGRA UI session at16 frames of boot-progress content (source_rgb stays ~8k px; R387 jumps to ~79k at frame5), the main render client keeps fi=0 (R387 fi nonzero), UnityPosterExten and ShortcutsTopHits crash, and ui-ready automation never arms. R389 reproduced the identical failure by accident (toml-inherited value, kept as a determinism control) and R390 restored e5-frameinfo=1 with byte-identical kernel/ICD/Data: desktop at79k px, complete Settings navigation, Home to a real desktop and successful Settings reopen (first pass of the r385-r387 boundary). Evidence docs/logs/gpu-r388-evidence.json, gpu-r389-evidence.json, gpu-r390-evidence.json and ui-r390/.
  • R385 captured the native Settings root, Notifications page and return to Settings. Physical touch cancelled the automated run after scrolling, so Home/reopen acceptance remained incomplete. R386 repeats the complete sequence with passive GPU timing.
  • R387 candidate restores native Mach FrameInfo transport: original slot336 true skips reference initialization for an unsupported shared queue. Pinned XNU returns Success without a send for a null reply port; the old list-node-as-reference fallback therefore did not establish delivery. The replacement walks only native live lists under their command gate, checks node+0x18 ports and uses two/six-word FrameInfo/vsync payloads. Host regression fails on the old sender and passes ASan/UBSan on the replacement; runtime FPS and complete navigation remain pending.
  • R384 fuses source pixel statistics into RGB565 conversion while preserving independent destination readback. Actual last-present work falls from roughly180ms to50ms; this excludes GPU and queue waits and does not establish smooth UI.
  • R377 candidate accelerates Bifrost register interference by tracking nonzero live nodes; exact constraints/affinity/live-state differential tests cover MOV, BLEND, dual destinations and partial masks through24000nodes (27-29x stage speedup). R376 delivered native Home and Settings taps but remained at two lock-screen frames while a helper pipeline compiled. Follow-up input mode4 adds frame-gated Settings row/Back/scroll/Home/reopen diagnostics; physical touch cancels and missing frames stop the sequence. All interaction outcomes remain pending.
  • R375 qualifies linear b3a8 XR10-to-RGB565 presentation: original backboardd completed two UI command buffers; both swaps accepted, copied and retired with zero failures. Actual plane layouts are 320x480, row1280 at0 and row320 at614400. FNV23d7ef95 frames show the original Chinese date/time lock screen and status bar. Host pixel/bounds ASan+UBSan and kernel build pass; Android A recovery, all three protected-window checks and complete Data backup pass. Physical panel and desktop interaction remain unqualified. See docs/logs/display-r375-b3a8-evidence.json.
  • R309 diagnostic makes the isolated Settings sequence wait for two actual CPU-presented frames before Home, then wait ten seconds before the tap. Synthetic mode3 requires its completed Home sequence and the native digitizer driver, not a physical sensor packet. Actual finger activity still defers/cancels the test. Other input modes and real hardware readiness remain unchanged. This corrects the R308 reproduction gate, not the Settings application itself.
  • R307 candidate expands opt-in deep task observations to 64 threads with truncation markers and reads the pinned libobjc runtimeLock before/after. Stack sampling now selects processes by name, restoring backboardd/lsd coverage after adding application targets. R306 shows SpringBoard main and HID threads waiting in Objective-C class lookup while Preferences waits for applicationInitializationContext; the lock owner and cause are not yet established.
  • R306 adds opt-in BGRA/source and RGB565/scanout summaries plus current Preferences/MobileSlideShow thread observations to isolate black Settings content. Kernel 814bccb2a646d48ac32307c078a530a8d761c8d96dbb63c53e1398d61d8255f3 builds with frozen source matching; host pixel/layout checks pass. Pre/postflash 37/37 and full payload readback pass; runtime and application content are pending. Original userland is unchanged.
  • R306 adds opt-in e5-ui-pixels1 source/destination pixel summaries at the existing retained BGRA copy path, logging dark/blank transitions and bounded periodic samples. Rejected format/composition changes are logged even late in a boot. Preferences and MobileSlideShow join the process/thread/dyld snapshots with exact same kernel ABI and static scratch storage. No pixel/fence/format acceptance or display behavior is changed; this distinguishes missing producer content from conversion and process startup failure.
  • R297 canvas mode reserves the validated firmware scanout for one diagnostic writer. The normal CPU presenter neither maps nor accepts compositor swaps in this mode, and e5-no-vc is mandatory. Readback-checked milestone images are canvas evidence only; no original UI copy/retirement, new DPU sequencing, vblank or physical power recovery is claimed. Host and device qualification are pending.
  • R286 integrates gated E5 HID startup and its archive before launchd. The accepted display/storage implementation and original Apple family instructions are preserved; new input runtime qualification is pending.
  • R284 hardware confirms at least 32 real BGRA-to-RGB565 copies and original retirements with zero copy failures; final swap API count 36. Pending masks change from 1:0 to 1:1 while surface identity remains valid. The user confirmed visible iOS UI. Android recovered and all 37 protected chunks passed without retained integrity/panic errors. Initial black-frame RLE capture is complete; later visible UI has no full capture. Exact desktop state and touch remain unqualified. See docs/logs/runtime-r284.json.
  • R283 hardware: IOPM callback sets requested=1 powered=1, then the CPU presenter accepts three real BGRA surfaces. Rechecking original mutable request masks in the pending queue rejects copying (reason6). R284 keeps the accepted full-frame plan and surface identity, rechecks actual format/backing bounds and retains the fence/retirement protocol. Pending masks are logged; R284 hardware qualification is recorded in docs/logs/runtime-r284.json.
  • R283 preparation registers two logical display IOPM states and retains the original setPowerState callback. R282 initial hardware log confirms requested=0 powered=1 after firmware reporting; the original submit gate still depends on requested. Two bounded RLE captures verify destination pixels with a complete-frame checksum for host PNG decoding; physical output remains separately unverified. R283 hardware confirms successful IOPM callback and three accepted BGRA surfaces, but no copied/retired frame; Android recovered and all 37 protected chunks passed. See docs/logs/runtime-r283.json.
  • R281 completed 8096 swap calls, unchanged panel CRC 11b0af80, no retained integrity/panic markers, Android recovered and protected windows passed 37/37. Presenter logging was suppressed; see docs/logs/runtime-r281.json.
  • R282 preparation reports retained firmware power through the hash-qualified original slot-354 state setter, preserving the client requested state. Bounded CPU presenter and state-transition events remain visible independently of e5-ca-state, including refusal geometry. R281 used the suppressed old helper; missing E5CPU rows in that run do not prove absence of validation/copy. R282 hardware: requested remains 0 and powered=1, 6663 swaps never reach CPU validation, no retained integrity/panic errors, Android recovery and 37/37 protected chunks pass; see docs/logs/runtime-r282.json.
  • R279 reaches the SpringBoard UIApplication run loop and real SwapBegin/SwapEnd, but original selector 3 fails with No Default FB Surface found; the user reports no SpringBoard pixels. R281 preparation connects hash-qualified original IOSurface getters, request validation, bounded BGRA-to-RGB565 copying and original current-swap retirement. SwapWait now uses the original completion state; host sanitizer checks cover padded rows, colors, alignment and rejected bounds. Default-surface absence is also seen on successful comparable ports and is not established as the sole blocker. No new display kernel is device-qualified yet.
  • R277 diagnostics: periodic CA inspection retains the last observed backboardd PID (with name/lifetime revalidation) instead of clearing it to zero. R275/R276 sampled the array before it was populated and otherwise lost every late display-state snapshot. Fix dyld_image_info stride from 168 to the ABI 24 bytes; the old image enumeration skipped six entries at a time.
  • R275 independent timer is live on eMMC-only: E5OBS/E5PROC/E5LIVE and E5DPU appear while PID 1 reads the cache. R275 CPU values exposed an older probe bug: cpu_time_eqos_stats at 0x4f8 is a task field, not a thread field. R276 samples real accumulated absolute ticks through recount_thread_time_mach while task_lock protects the member; prior E5TH cpu values are invalid.
  • R274 eMMC-only boots skip SD media publication, so the SD read-completion frame kick never runs and neither the runtime census nor the delivery/DPU timer starts. R275 starts an independent, serialized kernel-priority timer from display startup, samples processes/registry/panel/eMMC counters every approximately 32 seconds, and schedules the 60-second deep thread probe outside display delivery. The native family notification stays on the delivery worker. Census now includes launchd/logd/lsd/watchdogd and IOWatchdog registry presence. Host compilation is verified; R275 device acceptance is pending.

Completion criteria

  • A known frame appears with correct geometry and colors, then userland presentation and backlight controls work.

Current limits

  • Physical GPU panel observation, hardware-vblank timestamps, rotation, compressed surfaces and display/backlight power recovery remain unqualified.
  • R390 captures the complete Settings navigation including Home and reopen, but scroll has only near-top sampled frames and physical panel observation and hardware-vblank timing remain unqualified.

Next action

  • Settings functional acceptance is complete (r390); reduce interaction latency from the measured bottlenecks: synchronous queue submit waits (0.1-2.3s per SubmitWaitMany with 7-93 batches) and multi-second shader recompiles (19.4s observed inside a session; only weak_ref mem_cache exists).
  • Then prioritize battery/charging and shared Wi-Fi/Bluetooth power ownership.

Evidence

Input

implemented

Hardware and dependencies

  • sprd,sc9860-keypad matrix controller at 0x641B0000 (4 rows by 7 cols, KEY_STATUS event slots), AON GPIO bank 7 pin 15 volume-down (active low), PMIC EIC volume-up and power keys.

Configuration

  • e5-hid-ui-capture=10..60 selects periodic stable read-only scanout captures after120s, IDs300up; disabled bydefault and inCPUcanvasmode. It uses the existing single-pending off-workloop capture worker and does not inject touch.
  • e5-touch-irq=1 uses GPIO13 falling-edge detection and level-high sharedSPI40/INTID72, with a providerless IOInterruptEventSource on the owned HID workloop. The primary callback only acknowledges GPIO13 and queues work; no I2C in IRQ. e5-touch-irq-cycle=1 tests mask/drain/detach/reattach after300s and2sall-up quiet.
  • e5-touch-event-ms=1..16 selects the HID-workloop timer rearm interval only after strict GPIO event setup. Default/invalid values and non-event modes retain16ms. R299 uses2ms as a timing control; actual period includes handler work.
  • e5-touch-gpio-events=1 selects strict masked GPIO13 RIS polling after normal-mode restoration. Only an observed falling-edge latch authorizes a packet read; startup, held-low and recovery do not bypass the gate. Bounded event traces use mach_absolute_time ticks.
  • e5-touch-canvas=1 requires e5-hid=1 e5-touch=3 e5-no-vc, disables automatic tests/UI captures and ordinary compositor scanout writes. e5-touch-reset=2 requests a one-time startup sensor reset; e5-touch-raw=1 records bounded packet/IRQ-level evidence.
  • e5-hid=1 publishes a hash-qualified 207-slot IOHIDDevice subclass; e5-hid-test=1 enables one automatic Home press/release. e5-touch=1/2/3 selects resource, sensor-register and polled packet/HID stages. Runtime input remains unqualified until the isolated board trials pass.
  • UEFI: E5BootMenu installs a SimpleTextIn merge (keypad first, chained serial ConIn) on a vendor-path handle so the official UniXNU menu counts it.
  • XNU: e5-keys2.patch polls the matrix keypad in the serial keyboard path and feeds cons_cinput into /dev/console. AON gate 0x64900000 and AON GPIO 0x64170000 require set-mirror clock writes before peripheral access. The older fd230 e5-keys path is retained but is not the normal tty input path.

Procedure

  • Watch the menu stop its countdown on a physical key press and navigate with arrows and select.
  • Boot the seed and type through the serial console into the on-screen shell.

Development findings

  • R390 completes the full frame-gated sequence (open Settings, subpage row, Back, 17-release scroll, Home, reopen) with zero failures and all seven capture frames rendered; Home/reopen passes for the first time. The scroll phase executed with content changes but its sampled frame is near top, so scroll offset proof is weaker than r385.
  • R385 verifies100..115ms short taps, actual Notifications content and Back; real touch cancels automation after the scroll. R386 repeats the same coordinates and delays without mixed input.
  • R387 FPS diagnostics retain zero-update intervals for five seconds after input; reported interval and response-window rates remain distinct from physical refresh. Settings automation excludes capture/logging while held and preserves hardware-input cancellation.
  • R309 diagnostic makes the isolated Settings sequence wait for two actual CPU-presented frames before Home, then wait ten seconds before the tap. Synthetic mode3 requires its completed Home sequence and the native digitizer driver, not a physical sensor packet. Actual finger activity still defers/cancels the test. Other input modes and real hardware readiness remain unchanged. This corrects the R308 reproduction gate, not the Settings application itself.
  • R308 diagnostic adds a single synthetic Settings tap at the verified SE2 desktop coordinate 122,318 (mode3), after one Consumer Home pulse. Hardware activity retains cancellation priority; the test sends a down/up pair and no following gestures. The 64-thread/objc lock observer remains enabled to reproduce the application stall autonomously. No GPU or rendering change is included.
  • R305 manualPhone1x touch/Home/app-switcher/layout pass:601callbacks/599good,2coalesced,594event/5timer reads,0bad/recovery/foreign;294changes/23single taps/6drags/2dual releases/all-up/final0.426HID nofail/drop,originalclientdelivery,2433displayframes/30FNVcaptures,900s/fullreadback/allthree37/AndroidA pass. Allautomaticinputoff. Settings entersblack,post-settingsHome untested; currentB2D258C1 logscoveronlyearlystartup, so no uniquecause is claimed. See docs/input-se2-r305.md.
  • R305 frontend identity uses authentic iPhoneSE2 D79AP/iPhone12,8 product data from17.7/21H16 Apple DT. ProductIDc459fdb0eef455588dbc6f587c20a5372792bbf8 matches the current21H461 MobileGestalt table(row57,type0x80bb21e1). Artworkphone/subtype569/scale1 and chosendisplay-scale1 select native320x480 UIpoints. E5 board/UMS hardware, kernel/OS ABI,rootUUID/reservations and hardware compatibility suffixes remainfixed. A whitelisted retained-ADT edit proves unrelated property bytes unchanged; phone layout/runtime mustbe verified.
  • R304 preparation adds Home test5: two nativeConsumerMenu down/up pairs,100msheld plus150msgap, for app-switcher recovery. Physical matrix diagnostics have a separate256-sample budget so lateHome presses remain observable after hundreds of touchreports. R303 physicalMenu changes UItohome structure before the600sautomaticcontrol, butSettings becomesblack and singleHome/automaticHome cannotrestorevisiblehome; doubleHome operator test wasnotperformedbeforewatchdog. User900s authorization persists; nativeHome/pixels/appstate will be checked separately.
  • R303 preparation restores physicalMenu to native ConsumerMenu page0xC/usage0x40. The phone-own backboardd control logs explicitly classify this pair as Home; physicalCommand-H in R302 returns success atHID but has no lockscreen effect. Descriptor/nativefamily remain unchanged; ASan/UBSan key/chord/release tests pass. User explicitly authorizes900s nexttrial, with late600s singleConsumerHome control and periodicUIcaptures. R302 operator confirms originaltouch response and screenshots show lockscreen-to-wallpaperpicker via realtouch; Home/desktop remainunqualified.
  • R301 actual callback/transport pass:386 callbacks/readsgood,341deferred/45timer reads,158hardwarechanges/2dualcomplete/final0,operatortracks+clears,4FNVcanvasframes,allthree37/37 andAndroidA. Mask/drain/detach/reattach generation2 succeeds, but no post-reuse callback/input observation; complete reuse remainsINCOMPLETE. R300safe refusal was an invalidgroupvisibility check, not proof of securedSPI ownership. ActualR301 GICD_CTLRreads0x12 despite BSPalias request; successful callback delivery is the authority. See docs/input-gic-r300.md.
  • R302 preparation adds optional periodic UI snapshots, using the existing two-read stable RGB565/FNV capture worker. Capture queuing runs in the HID loop; frame emission stays off that loop. Busy requests retryafter1s and IDsadvanceonlyonqueued requests. This allows readiness and actualfinger actions to be observed before the old550s-onlydeadline. CurrentR301 frozenkernel keepscanvas/GIC isolation; originalUI acceptance will follow.
  • R300 safely refuses the masked parent at readbacka000 before enablingGPIO13; detach/mask/drain return0. Linux irq-gic-v3.c states IGROUPR writes/readback are only relevant in single-security-state mode. R301 records GICD_CTLR/DS visibility, requires group1 only when visible, and verifies nonsecure ISENABLER access when enabling. ArmCpu now enables both Group1 aliases in the dual-state nonsecure view, matching BSP0x13; single-state view retains0x12 and does not additionally enableGroup0. No DS/security-state change is made.
  • R300 preparation adds ArmCpu masked attach, owner-qualified enable/mask and gate-based detach/drain through an isolated local module override. A closed atomic gate excludes new context users; timeouts retain closing ownership and resources. The GPIO callback censuses all16shared banks, acknowledges onlyGPIO13, and contains foreign-source conflicts without clearing their events. Deferred work consumes callback counts, records coalescing, and retains unread callback tokens through500msI2Crecovery; the16ms timer provides retry and status checks but no idle packet reads. Optional idle lifecycle cycle validates reuse. Host and actual callback/UI qualification remain separate.
  • R299 completes the isolated2ms GPIO event/CPUcanvas trial:301 valid packets,zero errors/recoveries;274433GPIO polls/301reads/274132skips.65 sampled timestamp orders and actual2ms setting pass.232hardware changes include3single drags/1dual sequence with complete all-up,final0 and no error-terminated sequence. Operator confirms tracking and ring removal.244HID reports/fail0/drop0,frames200..203 original checksums,full payloadreadback/allthree37/37/AndroidA/no retained panic pass. Optional600s canvas frame204 absent; GIC/originalSpringBoard actions remain separate. See docs/logs/input-gpio-cadence-r299.json.
  • Source audit of selected ArmCpu9066b53bbc82e92a98ec44a11dd50fd13eaf1a1b finds armcpu_spi_attach immediately unmasks the parent and armcpu_spi_count observes callbacks, but no mask/detach/inflight-wait API exists. The primary dispatcher calls handler/context before EOIR. This lifecycle interface must be added before an owned GPIO callback can release resources safely. Live FDT gpio@170000 interrupts0/0x28/4 confirms sharedSPI40/INTID72 level-high, independently of GPIO13 falling-edge selection.
  • R299 preparation adds a bounded1..16ms event cadence option without changing I2C, decoder or one-read-per-consumed-latch semantics. R298 live evidence shows ordered traces and consistent masked GPIO accounting with idle read suppression, but two malformed packets each coincide with pending_after1. This correlation does not establish a race;2ms is a subsequent timing control. R298 Android A, three37/37 windows and five original-checksum canvas frames pass;276 hardware changes,two dual releases/final0 are observed, while transport remainsFAIL.
  • R298 preparation consumes GPIO13 RIS with pin-only W1C before I2C, preserving edges arriving during the transaction. IE remains masked; no GIC callback is installed. Timer/edge/read/skip/low-without-edge counters and bounded observed/start/end timestamps distinguish read gating from packet validity. RIS can coalesce edges and observation time is not the physical edge timestamp. Host/build and device evidence are recorded separately in docs/input-gpio-events-r298.md.
  • R297 logs contain 1088 hardware contact changes, 12 complete two-contact releases and final active0. The operator confirms five targets, drag, two rings, ring removal and Home indicator. However 18 malformed packets/18 recoveries remain; the GPIO census reports other_enabled0/other_pending0 across16 banks at startup only. See the e5-xnu R297 archive; this is not error-free transport or GIC qualification.
  • R297 preparation adds an opt-in exclusive CPU touch canvas, a bounded raw-frame/key queue, stable-ID trails and active-ring removal. Errors and queue loss cancel old paths; worker stop waits before releasing the validated 0xbe000000/307200-byte mapping. Readback-qualified milestone frames use the existing encoder. It is raw sensor-to-scanout evidence, not UIKit acceptance; host and device qualification are pending.
  • R297 e5-touch-reset=2 tests one startup GPIO14 high10ms/low10ms/high60ms reset after pin qualification, matching the observed Android resume path for flash-backed3536. Mode1 remains conditional after4 failed identification reads. Firmware/configuration/normal-mode checks remain mandatory; no firmware programming. e5-touch-raw=1 has bounded separate two-contact packet budgets and before/after GPIO13 samples. The I2C algorithm and strict decoder stay unchanged.
  • The R296 operator repeat of device5bacf9d FAILS: last periodic40961 reads/40955good,6 packet-format rejects/6 recoveries,311 HID reports without reported failure/drop,61 retired frames. Keys and bottom-edge swipe do not unlock; hardware logs max1 contact. Later Android control has two complete simultaneous slot0/1 sequences, so dual touch is possible on the phone. Original backboardd delivers paired physical touch events to SpringBoard, but keyboardFocus buffering times out and the later automatic keys have no destination. See the linked e5-xnu reflash evidence; the earlier idle R296 PASS is separate.
  • R296 device5bacf9d terminal-FIFO fix passes:40961 valid packets,bad0,recoveries0;18 HID reports/fail0/drop0,1366 retired displayframes;all100/101/102 captures and all three37/37 windows/fullpayloadreadback pass;AndroidA/no retained panic. Snapshot102 black after Photos tap:app contents remain visually unqualified;R295 UIKit action evidence stands. Later IRQ-census f4d9e13 is host-built only. See docs/logs/input-runtime-r296.json.
  • Post-R296-freeze e5-touch-irq-probe1 reads all16 GPIO banks after qualified clock/map setup. Records IE/RIS/MIS/IS/IBE/IEV and nonowned-bank masks; no interrupt-register write or parent attachment. SPI40/INTID72 is shared, so GPIO13 masking alone cannot contain an unrelated pending bank. IC is W1C, never saved/restored. ArmCpu parent mask/detach and callback lifetime remain prerequisites; board census pending.
  • Recovered R295 device unified logs from read-only R296 Data backup prove UIKit Direct Touch195,425 hits Photos Dock, launches com.apple.mobileslideshow and completes home-screen-to-application transition. Backboardd touchstats down1/up1/destinationSpringBoard. See docs/logs/input-ui-r295.json. Automatic UI action is verified; missing102/physical actuation remain unqualified, and transport timeout still makes overall R295FAIL.
  • R295 digitizer UC belongs to backboardd pid57: open1, enqueued2, dropped0, type11, queue head=tail0x1e0. DisplayIntegrated/Built-In/HIDServiceSupport true. Queue consumption is proven; Photos UI is unqualified because capture stops around621s before snapshot102. Original runner incomplete; independent Android A/postboot37PASS on resume. At least24 frames/no retained panic, one recurring FIFO timeout0x310420. R296 tests the qualified terminal-tail fix with unchanged HID/tap control and complete capture.
  • Post-R295 preparation: I2C terminal-tail handling retains first-interrupt readiness qualification, but drains an idle nonempty final FIFO after a previously qualified read even if its final IRQ was lost. Host race fixture reproduces old Timeout/received16/status0x310420 and passes after correction; initial unqualified FIFO, terminal NACK/short read, final-byte NACK, and existing protocol ASan/UBSan cases pass. R294 race is plausible, not device-established; frozen R295 still uses the old bus code and new transport requires a later trial.
  • R295 preparation extends the existing off-workloop public registry probe to event-service properties and child user clients. DebugState EnqueueEventCount, DroppedEventCount, LastEventType/Time and EventQueue XML distinguish original report processing from client delivery/consumption. IOUserClientCreator and per-client isOpen identify real subscribers; no original method/vtable or report descriptor changes. Touch-test2 taps only the fully visible Photos Dock at195,425, then leaves the UI unchanged until capture; mode1 keeps the gesture suite. Runtime evidence pending.
  • R294 native IOHIDEventDriver attaches and processes the separate digitizer reports (two transducers, grouped, LastReportTime present). All16 synthetic phases/32 HID reports finish without report errors or active contacts; Settings does not open in checked snapshots101/102. One I2C packet timeout status0x310420 means overall automated FAIL.33 display frames, full payload readback and pre/postflash/postboot37/37 pass; fresh Android A and no retained panic. See docs/logs/input-runtime-r294.json.
  • Split report3 submission now enters the digitizer instance's own workloop with runAction, matching the audited Sheng caller invariant. This is synchronization hardening, not a proven explanation for R292 UI failure; the native-family handleReport internal locking was not established from available sources. No transport/descriptor semantics change.
  • R293 split-service metadata identifies a waiting AppleUserHIDEventDriver proxy with no Digitizer/DebugState, followed by AppleUserHIDDrivers server timeout. The original kernelcache has no ordinary native TouchScreen13/4 personality; the SafeBoot broad13 case is vendor-specific. New touch-native1 installs an owned native IOHIDEventDriver/IOHIDFamily catalogue match for13/4, score60000 before publication. Original module instructions/vtables are unchanged. Discovery now requires DebugState instead of accepting an unstarted subclass; public probe logs actual class names. This matching change still requires its isolated device trial.
  • R292 additionally records6 malformed packet rejections (Count/Inconsistent) during physical contacts, with successful I2C transfers; prior raw bytes were not retained, so no torn-read/palm/decoder diagnosis is established. The next transport log includes all18 raw bytes and pre-abort status for invalid packets, retaining strict validation and existing lift/recovery behavior. UI routing remains the priority of the split-service trial.
  • User R292 confirms desktop entry but no touch UI response although hardware move/lift packets and HID reports succeed. Snapshot100 is black at440s and101/102 showlockscreen, so450s is not a qualified UI-ready deadline. Next touch-split1 uses a second instance of the same audited207-slot class for a touch-only primaryDigitizer service, independent workloop and public properties; main service retains keys and owns the sole I2C poller. Services publish afterstart; an opt-in off-gate public Keyboard/Digitizer/DebugState probe records grouping/LastReportTime without private ABI hooks. Pre-suite hardware activity defers untilallup+2squiet; in-suite input cancels. Sensor readiness resets on transfer recovery. The next bounded trial uses550s keyboard/600s touch/640s snapshot.
  • R2911301201 reads3536 boot0x0240/config0x0c0e0dd0/chip0x2d5c,204-byte config SHA19cc5ee07b370300680e4136c0ded6225cc21ee405716da29fc8a3a196b0a820 with zero16-bit checksum and matching slave0x2e. Initial padsIRQ/reset30,SDA/SCL0 are already correct; no mux field changes or conditional reset pulse occur. Pin clock/reset GPIO initialization/settle differ from R290, so their individual effects are not isolated. Sensor packets have no errors/recoveries and all16 synthetic touch phases complete with no active contact; physical finger/UI effect is unqualified. Next code validates config checksum/address before packets, reads every16ms to avoid missed GPIO pulses, targets Settings with synthetic touch and takes snapshot102. Physical Menu candidate changes to Command-H with release/chord coverage; Consumer Menu diagnostic remains independent. Source header added to ABI audit inputs.
  • Before optional pad-mode changes, GPIO14's high output latch/direction is prepared, then the pad is connected and given60ms to settle. This avoids a stale low-latch reset glitch; the explicit reset test remains one pulse after4 failed identity reads. Other pad fields and the IIC matrix route are preserved.
  • User R290 test confirms physical arrows move UI focus; Home remains faulty and touch is prioritized. R290 completes199 frames,160 matrix transitions,336 reports and a live observer, with no failures/drops/panic; AndroidA/post37pass. Fresh Android probe identifies3536/config0x0c0e0dd0 and resets on probe/resume before idle suspend. Opt-in touch-pinmux1 selects proven GPIO13/14 func4, mode2 also tests inferred dedicated SDA3/SCL3 func1, with clock-before-access, bit-field-only RMW/readback and stop restoration. No matrix-route or firmware programming.
  • Early R2903e6a554 still reports NACK with a live observer, so the IRQ-qualified ACK correction is not the only blocker. Next code retains pre-abort status/read-or-write phase/time, adds opt-in read-only pads/IIC-route/line probes with independently derived pinctrl PA0x642e0000 and pin-eb EB1 bit11, and permits one vendor-style reset only after4 failed reads and a GPIO14 func4 pad qualification. No I2C pad-mux write or firmware programming; RAM-only runtime code is preserved by default.
  • R289 completes Consumer Menu/Command-H/Return/Command-H with16 accepted reports and a live independent observer; checked snapshots show Chinese lock screen to desktop and at least35 frames retire. This proves the sequence effect, not physical Menu or the effective single action. Touch mode2 identity fails on NACK; Android A and postboot37/37 pass. Single-action test modes2=Command-H and3=Return are now prepared with before/after captures; R2903e6a554 remains frozen for the isolated ACK fix.
  • R289 early observer/options appear and the HID tick progresses at stage10; touch A3 returns NACK. BSP ACK is checked in IRQ context, whereas the first adapter checked RX_ACK before IRQ. The correction gates NACK by I2C_INT and adds a transient-PendingAck host negative control. Registry walking now stops after first event-driver attachment and queue bursts are limited to8 transitions; stop and producer-side monitor fields distinguish termination from stalls. User physical tests still report no key/touch UI response; current e5-touch2 is identity-only.
  • R288: corrected AP-AHB gate0x20000000 changes89->2089, I2C3 version900, IRQ/reset1. Thirty-six matrix reports complete with zero HID failures/drops, but HID workloop telemetry stops and timed Home/captures do not occur. The user reports no key/touch UI response; touch mode1 was resource-only. R289 adds an independent stage/options observer and moves scanout capture off the HID workloop before sensor packet tests.
  • R287 stable read-only scanout captures100/101 pass full checksums and show a Chinese lock screen before/after Consumer Menu. Clock text is clipped at the narrow viewport edges; no desktop or unlock transition is qualified. The next diagnostic sequence tests Consumer Menu, Command-H, Return and Command-H separately while preserving normal physical key mappings. Source-driver gate translation is now independently derived from real FDT ranges, including a negative control rejecting0x20003000 before device access.
  • R287 resources: the obsolete gate base0x20003000 read zero before/after the set write. The driver safely refused controller access, without panic. FDT apahb-gate reg is address0,size0x3000 under syscon@20000000; corrected base0x20000000/set0x20001000. R288 will qualify this before sensor transactions. A gated synthetic tap/drag/two-contact HID test is also prepared; ambiguous report completions retain possibly-down IDs for explicit future lifts.
  • R286 hardware: hash-qualified original IOHIDDevice parent size208 and owned size288; native IOHIDEventDriver attached. Synthetic Home down/up reports both return success, with four accepted keyboard/consumer reports, zero drops/failures and held state0. Fifteen real display frames complete; no retained integrity/panic markers; Android A and protected windows37/37 recover. Physical keys and desktop effect remain separately unqualified.
  • Next input revision separates keyboard and Consumer Control Application collections, maps raw TLSC IDs0/1 to HID IDs1/2, qualifies big-endian 16-bit sensor addressing and records firmware/config identity through temporary direct mode with mandatory normal-mode restoration. A gated read-only before/after scanout capture verifies autonomous UI changes. No firmware-flash update or new DPU writes are introduced.
  • E5 KEY_MENU at matrix row3/col0 maps to Consumer Menu usage0x40 (iOS Home); Back maps to keyboard Escape, Select/Phone to Return, arrows and digits to keyboard usages. The existing serial scanner is the sole owner of keypad INT_RAW and feeds a 64-entry transition queue, preserving releases and same-scan fast taps. GPIO127 is explicitly configured as an input before volume-down sampling.
  • Host ASan/UBSan checks pass for wiring/Home/chords/fast taps, contact bounds/stable IDs/lifts, and polled I2C repeated-start/FIFO/NACK/timeout recovery. Original IOHIDFamily text, constructor/descriptor anchors and 207-slot ABI are hash-qualified against the original 21H461 kernelcache. See docs/logs/input-host-audit.json and input-source-inventory.json.
  • The actual TLSC6x bus PA is 0x22270000 after AP-AHB translation. Its AP-AHB gate is bit13 at0x20000000 (set mirror0x20001000). AP-I2C3 mux is at0x200100a0 and divider at0x2001009c; AON GPIO needs its clock first. The first implementation preserves working pinmux/reset/firmware, uses 400kHz register transfers with 10ms bounds and releases all HID contacts on errors. GPIO IRQ is initially polled; interrupt routing and firmware/config identity still need qualification.
  • The keypad block reads sane values in XNU (E5KINIT 000f7f030000000077777777: CTRL row/column enables correct, KEY_STATUS idle pattern). The caller is the implemented e5-keys2 serial poll hook.
  • Fixed-v8 controls: original kernel 242 fails (#35); changing only two address-forming instructions succeeds (#36), and source-built 244 succeeds (#37). An e5-keys token with kernel 241 also boots (#38). Earlier claims of general binary-layout or token-content fragility are withdrawn.
  • Flash #41 verifies real console fd0, TIOCGETA=0 and ICANON=1, four commands exiting 0 and a blocking canonical read. Earlier CONSOLE-FD 2 was an unhandled ENOENT from a missing /dev directory. Physical key and serial command entry still require separate on-device acceptance.
  • Stock Android enumerates tlsc6x_touch in addition to sprd-keypad and gpio-keys; the claim that this board has no touch panel is withdrawn. Touch hardware binding and XNU HID support remain pending.

Completion criteria

  • Touches map to the display, simultaneous contacts keep stable IDs, release events arrive, and buttons produce expected HID reports.

Current limits

  • Physical transport and Home/switcher evidence from R301/R305 remains tied to its retained configuration; suspend/resume and panel power recovery remain separate.
  • The current Settings icon is200,186 and Notifications rowY389. R385 subpage/Back are verified; full isolated Home/reopen and sustained smoothness await R386/R387.

Next action

  • Capture a visibly offset scroll frame (denser scroll-phase capture or hold position) to close the last visual gap; physical-input cancellation stays retained.
  • Validate power-key/wake behavior during PMIC bring-up.

Evidence

GPU

implemented

Hardware and dependencies

  • Rongyue E5 Mali-G57 Valhall v9 Job Manager, GPU_ID 0x90910010, shader mask 0x5, VA48/PA40, AS 0xff, JS 0x7; stock Android kbase r41p0 UK 11.38.

Configuration

  • e5-mali-g1=1 e5-metal=1 enables the native kernel/Metal stack; the original compositor renders with GPU, followed by CPU RGB565 conversion.
  • ABI1.7 accepts both64/256-handle SubmitMany packets and offers combined SubmitMany/WaitAtom with unchanged legacy APIs. PANVK_E5_SEPARATE_WAIT=1 selects the old userspace round-trip path for measurements.
  • Imported shared memory retains original CPU cache attributes. Every pinned imported range is cleaned before submission and writable imports are cleaned/invalidated on actual terminal retirement.
  • e5-gpu-perf=1 samples the first8 and every64th job. Its out-of-object timing table preserves the IOService size limit; the diagnostic is disabled by default.

Procedure

  • Freeze source and kernel hashes, run the e5-xnu e5-uefi-cycle pipeline with 900-second watchdog, verify all 37 partition chunks before flash, after flash and after return, and capture a verified Data-after backup.
  • Require actual job completion, zero unexpected errors, and independent full-frame pixel/hash comparison before accepting rendering; repeat reused descriptors before testing imported scanout.

Development findings

  • R388 adds ABI1.7 timing instrumentation to the first working UI session (r390): E5SHADERPERF (panvk_compile_shaders) and E5QUEUEPERF (gpu_queue_submit) sample first8/powers2/slow calls; E5GPUPERF samples submit cache, true GPU/IRQ completion, observation latency and retire cache per atom. Measured r390: single shader compiles reach 19.4s inside the session (r388 boot saw 15.5/15.1/9.9s; r389 max4.9s - a distribution, per-run order varies) while the runtime holds only a weak_ref mem_cache (disk cache disabled in this Meson build), so pipeline destruction forces recompiles; queue submits with 43-93 batches take 0.6-2.3s under the single synchronous SubmitWaitMany call while individual atoms execute in ~100us with ~4-6ms kernel overhead. These two paths, not the RGB565 copy, dominate interaction latency. RGB565 paired scalar access and the FrameInfo count1 advertisement are retained as verified behavior.
  • R386 sampled per-job cache clean, GPU/IRQ completion, completion observation and retirement sync separately. First-eight startup jobs have larger scheduling delays; later sampled cache stages are about2..5ms each, hardware/IRQ time about0.2..29ms and observation delay near zero. Do not generalize startup latency to steady-state FPS.
  • ABI1.7 adds selector16 combining the existing SubmitMany and bounded500ms WaitAtom without a second IOKit round trip. The driver releases its state lock before waiting and reuses the exact cache publication/pin/retirement path; active timeout/fault is not DONE. Mesa uses the new call only on minor>=7, with the legacy two-call path retained. Actual-source ASan/UBSan tests cover both paths and failed/pending completion refusal; the new signed ICD/kernel await R387.
  • Source inputs are pinned with SHA256-checked ordered Mesa, MaliMetal, Metal2Vulkan and ArmCpu patches. The ArmCpu lifetime dependency remains device-carried rather than pushed to the shared upstream.
  • R313-R336 establish device identity, MMU/BO operations and JM write-value readback. R332 and R337 clear tests produce exact full-frame RGB565 results.
  • R337 preserves the complete GPU-written TCTX polygon value. Fullscreen VS/FS rendering passes with mismatch=0 and FNV cf06c8848298b325, with two submitted/completed jobs and zero errors. The second variant faults at reused job descriptors; their GPU-written 16-byte header prefixes were not reset.
  • R339 fixes the analytic fragment shader: integer upper-left pixel coordinates need +0.5 pixel centres, no vertical flip, the corrected edge direction and floating-point comparisons. Fullscreen and triangle each pass eight alternating repetitions; every triangle has mismatch=0, inside=39810 and FNV 65013b65352f3955. Clear also matches exactly, with zero GPU failures, faults, timeouts or resets.
  • R339 runner rc0, full payload readback and all three 37/37 checks pass. Android A returns with a new boot ID and unchanged CID/sector count. The verified 4GiB Data-after hash 8075a4a5... matches R338. Full frozen evidence is in docs/logs/gpu-r339-evidence.json.
  • R340 compiles and passes all offscreen controls (34 submitted/completed, zero faults/timeouts/resets/MMU faults) but rejects scanout before submission. Import accounting and teardown return to zero. Host reproduction proves entryPhys retained NX bit54: the valid leaf 00400000be000743 decoded as physical 00400000be000000 rather than be000000.
  • R341 fixes NX leaf physical-address decoding and passes 75/75 imported pages at PA be000000. All 16 direct scanout frames match the independent host reference byte-for-byte, with two stable reads and zero CPU presenter copies. Total jobs are 66 submitted/completed; faults/timeouts/resets/MMU faults and active/mapped/BO resources are zero. Full 900-second run, runner rc0, all three 37/37 checks and fresh Android A return pass; the complete 4GiB Data-after SHA256 matches R340 and the latest-three backup retention pass completes successfully.
  • R343/R344 exposed a probe plist but iOS 21H461 loaded its signed precomputed launchd cache; no user-probe entry was captured. R344 USB disconnected before the end of the window and automatic Android return timed out (runner rc1). The user reconnected/powered the phone; harvest, postboot 37/37, Android A identity and a complete SHA256-verified Data backup then closed recovery without relabeling the automatic run PASS.
  • R345 verifies genuine userspace G1 ABI 1.1: a normal LC_MAIN program ran as PID1 before launchd, opened E5MaliG1, queried raw properties, mapped 64 BOs, rejected invalid cache/65-handle requests, verified last-handle pinning and completed eight RGB565 GPU clears with zero mismatches/FNV e2c50ed21b7e2325. Submitted/completed=8/8; faults/timeouts/resets/MMU faults and final active/mapped/BO counters=0. It execed original launchd in place. Runner rc0, three 37/37, Android A and complete Data SHA256 4837f188ca13910827c75a66130993cfa89f977c6d0b3d41a364878743c52201 pass. PanVK, Metal, original SpringBoard GPU rendering and physical observation remain unqualified. See docs/logs/gpu-r345-evidence.json.
  • R345 restores the official 21H461 Skip Setup policy on media-direct eMMC Data using a clone of the fresh R344 backup. A 364544-byte delta passed independent command replay, full installed 4GiB readback and adjacent-sector/GPT/37-window checks. All five policy dictionaries and zh-Hans/zh_CN persist after boot with mobile UID501; Setup has no captured process census, while SpringBoard is active. containers/Temp was verified before install and absent after boot. Iconservices crash markers remain; no visible home-screen claim is made.
  • R346 prepares explicit mode3 startup for a signed probe and native PanVK library staged under owned Data. The kernel checks the staged probe is a bounded regular file before ordinary exec; mode2 retains the verified embedded ABI probe. Vulkan execution and UI qualification remain pending this new build.
  • R346 reached its signed staged PID1 probe but dlopen of the PanVK ICD failed with file-system sandbox blocked mmap; no Vulkan GPU job is qualified. The probe handed off to original launchd. R347 prepares e5-gpu-map=1: diagnostic reporting and a development-only MAC mapping grant for that exact ICD path, qualified eMMC Data, PID1 named e5vk-probe, private mappings and no RWX. Descriptor permissions and later VM/code-signature validation remain. R347 mapped the ICD executable segment after the original MAC EPERM, opened the genuine E5MaliG1 client, then trapped on MRS CTR_EL0 in util_cache_granularity during physical-device properties. No GPU job ran. PC and LR match the signed ICD instruction and caller at image offsets 0x19b538 and 0x70ee8. The fatal handler retained PID1 until the watchdog; automatic return, runner0, three 37/37 checks and a complete 4GiB backup passed. See docs/logs/gpu-r347-evidence.json. R348 uses the kernel ABI 64-byte cache granularity in property queries and deferred-sync coalescing. Native Vulkan compute (1024 integers) and a 320x480 RGBA triangle both pass with zero mismatches; submitted/completed4/4, faults/timeouts/resets/MMU faults0, both client resources0. The probe execs original launchd and SpringBoard/backboardd remain active. Runner0, three37/37 and a complete Data backup pass (SHA256 2d976207567ffdcab7ab26e24e013d155fa7d884bd556816e72bdb0ee3ab2209). See docs/logs/gpu-r348-evidence.json. Metal and original UI GPU composition remain unqualified. A separate Metal candidate adds opt-in mode4 for e5metal-probe and map2 for that PID1 probe to load only its signed plugin and ICD. Mode3/map1 behavior remains intact. Authored Metal compute and triangle passed on host Apple M5 (zero mismatches, triangle FNV 417503084e2ac225); iOS shader translation passes within 20s/500MiB. The separate kernel candidate compiles as SHA256 60ae7868ccba4a8fc874536b80d2c72817f99368631c274de6d38897bf6aa8b2. Its Metal ICD also exports vkGetInstanceProcAddr for the Rust bridge direct-loader contract. R349 dedicated Metal compute/graphics pass on E5 with exact reference pixels, five jobs completed and zero final resources; full round closeout is in progress. Original UI GPU composition remains unqualified. The next candidate adds ABI1.2 ImportUser for page-aligned writable ranges in the retained owner task, non-executable GPU mappings, kernel cache maintenance, and DMA-safe unpinning. PanVK advertises host import only for ABI1.2 and accepts only non-coherent cached memory types. R350 rejected all six invalid/read-only imports and retained ordinary Metal rendering, but the runtime refused noncoherent no-copy memory before a GPU import. The ABI1.3 candidate holds write-combining through XNU pmap per imported physical page, with reference counts across clients and restoration on last release and exposes coherent host imports only on that ABI; physical GPU IO coherence remains false. R351 confirmed alias reference counts and exact per-page cache restoration, but the first no-copy dispatch was rejected before hardware submission: the legacy submit validator omitted userMd-backed BOs. The next kernel recognizes every descriptor backing in submission and free-slot selection; it reuses the signed R351 programs and a 600-second focused window. R352 passes both CPU-to-GPU-to-CPU no-copy rounds, survives releasing one of two client references, and renders a real IOSurface with zero pixel mismatches. All nine jobs complete with zero faults/timeouts/resets/MMU faults; cache restoration mismatches and remaining imported-page references are zero. Original launchd handoff succeeds. Full round closeout and original UI qualification are tracked separately. R354 creates the original backboardd Metal device after using its sandbox-approved com.apple.backboardd cache namespace, then rejects QuartzCore framebuffer fetch. R355 enables E5 v9 color rasterization ordering and native LD_TILE, but its ordered uint recurrence fails with a zero flat varying; no hardware faults or leaked resources occur. Host compilation preserves the uint output, secondary vertex program and flat fragment load with matching location offsets and no FP32 flush. R356 tests exact generic varying packet strides against constant and vertex-id controls plus independent uint and UNORM fetch recurrences; all four Apple M5 references pass. R356 all four rendering cases fail: constant and vertex-id varyings read as zero, uint fetch follows the zero-step recurrence, and UNORM readback retains initialization bytes. All 16 jobs complete without GPU errors or final resources; recovery and full Data backup pass. R357 corrects uncached/coherent ordinary allocations by owning anonymous CPU memory and importing it through the proven ABI1.3 page policy. It combines real compute/render checks with opt-in post-VTC heap/packet readback; R357 ordinary coherent allocation and compute pass; UNORM now reads rendered data rather than its initialization bytes. GPU heap readback finds all 48 sentinel writes at the packet varying area, while fragment inputs still read zero. All 17 jobs complete and final resources/cache references are zero. R358 uses typed descriptor-based v9 fragment varying loads, explicit packet offset1024 and strides, and reserves the varying descriptor prefix before sampler/dynamic-buffer lowering. Host compilation emits LD_VAR_FLAT_IMM.u32; allocation lifetime checks still pass without GPU submission. R358 passes genuine compute, both varying controls and the ordered uint/UNORM framebuffer-fetch recurrences with exact Apple M5 reference hashes and zero pixel mismatches. All17 jobs complete with zero GPU errors and final resources/cache references. The typed descriptor path resolves the fragment varying failure. R359 original SpringBoard/backboardd rendering is queued after automatic R358 recovery, using the identical signed user stack and original mmap policy. R359 fails before opening the native device: its probe-only trust module omitted the different system-plugin CodeDirectory, while the system bundle still contained the old ICD. R360 stages the actual system-extensions bundle, preserves the working R354 frontend text, embeds the exact R358-qualified ICD, re-signs resources and trusts both final system-path binaries. R360 now opens the native device from original backboardd with the correct cache namespace and no signature rejection, then endEncoding reports OutOfMemory(Device) before GPU submission. R361 adds bounded failure-only Vulkan file/line, E5 allocation size/flags and kernel-return/state diagnostics; allocation limits and successful rendering paths are unchanged. A host forced oversized allocation remains a typed failure and leaves zero resources. R361 reports selector1 kIOReturnNoSpace for a64KiB flags0x41 BO with only44,400,640 bytes mapped: the64-handle table is full. E5 command buffers were eagerly allocating a GPU descriptor BO even for CPU-queued secondary buffers. R362 defers descriptor backing until actual use. A host160-idle-command-buffer regression consumes7->64 BOs before the fix and remains7 after; recording a secondary remains CPU-only and primary replay allocates backing successfully. No allocation limits are raised. R362 removes the BO exhaustion and original backboardd completes at least64 real GPU jobs, but captured UI frames stay black because IOSurface imports fail. R363 identifies the first page as managed RAM with VM_WIMG_RT0x17; raw and physical segments agree. R364 accepts only managed RT pages as already uncached, preserves their cache mode throughout shared references, and leaves ordinary cached-page WC conversion/restoration intact. Actual-namespace host regressions cover RT aliases, cached aliases, unmanaged/unknown rejection and mixed rollback. R364 accepts the actual managed RT IOSurface pages and reaches75 completed jobs, but later16KiB shader allocation exhausts64 BO handles at69,537,792 mapped bytes; original UI remains black. R365 candidate uses256KiB persistent shader/descriptor slabs and ABI1.4 explicit GPU RO+NX host imports, retrying only KERN_PROTECTION_FAILURE. Read-only CPU cache attributes stay unchanged and the kernel cleans sources before submission. Authored two-round Metal readback without explicit CPU flushing passes on Apple M5; kernel and iOS userspace build successfully. R365 hardware passes both readonly Metal reads and cleanup with2/2jobs and zero errors. Its isolated write probe gets the correct PERM_FAULT_3 at the targetVA, but recovery fails: irqPrimary clears RESET_COMPLETED while resetEngine polls RAWSTAT with that IRQ enabled, causing three lost completion events per attempt and a reset storm. R366 masks GPU IRQs while polling reset completion. An extracted actual-code host regression reproduces failure before and succeeds after. R366 hardware fully passes: two readonly Metal read rounds have zero mismatches and zero positive-path faults; deliberate write gets exactly one PERM_FAULT_3 at the target, one reset, zero timeouts, unchanged source bytes after invalidation and zero final BO/mapping/atom/cache resources. Original launchd handoff succeeds. R367 original backboardd imports its readonly16KiB source successfully at handle33 and no longer reports the prior BO exhaustion, but captured frames remain black. It requests HOST_ALLOCATION_EXT images and PanVK rejects that handle in image-format queries despite the implemented host-memory backing. R368 enables import-only support for single-plane2D linear color images, including an explicit linear modifier, while rejecting unsupported layouts/depth/multiplane and requiring the enabled native host-import capability. An actual-function host regression reproduces the missing capability before and passes supported/rejection cases after. R368 demonstrates actual shared-image binding of two256x1 RGBA16F images, but original UI still stops at75GPU submissions and black frames. SpringBoard fatally faults in release_shmem_bitmap at LDAXR on the shared bitmap reference count, then restarts. The bitmap was allocated as ordinary shared RAM; writable GPU imports currently retype its whole page to non-cacheable memory. R369 candidate preserves all existing CPU cache types, cleans imported pages before GPU submission, and cleans/invalidates writable imports on retirement before releasing pins. This preserves CPU-owned metadata cache lines modified while the GPU accesses other lines. ABI1.5 and a focused no-copy bitmap-header atomic/GPU-write probe are added; Apple M5 reference passes both rounds, untouched guards and IOSurface clear. Actual-namespace host cache policy tests pass. R369 hardware passes the named-shared purgeable alias test: atomic header access stays legal while imported, counts10001/20001 are exact, both GPU read/write rounds have zero mismatches and unchanged guards without explicit CPU flushing. A direct imported linear64x64BGRA IOSurface GPUclear has the exact M5 hashd0530839d87ea325. All3jobs complete with zero GPU/MMU faults/timeouts/resets and zero final resources/cache references. R369 completes900s without the earlier SpringBoard crash, but no application command-buffer completion or visible UI appears. R370 reduces alignment only for uncompressed single-plane linear host images from4096 to the underlying64-byte slice alignment, keeping imported memory itself page-aligned. Its actual GPU texture-read test directly binds a64x16RGBA image at offset64 behind an atomic bitmap header and passes every pixel, guard and reference count. All5jobs complete with zero errors/final resources; IOSurface direct GPUclear still matches the reference. Detailed layout failures and32-thread backboardd stacks with private image bases support diagnosis of the remaining UI wait. R370 full900s still has no visible UI or application completion, while SpringBoard stays alive. Full backboardd stacks locate the display thread in waitUntilScheduled, waiting for graphics pipeline compilation; workers are in SPIR-V/NIR preprocessing and allocator waits, while the encoding worker is idle. The runtime enables compiler and completion QoS requests only on macOS, leaving iOS as no-ops. R371 extends those existing Darwin paths to iOS and adds bounded driver compile QoS/effective-priority and wall/CPU timing. The actual-function iOS codegen regression has no priority calls before and emits them after; iOS release bridge/frontend builds and target17-system import checks pass. R371 new thread-statistics diagnostics triggered SANDBOX25/code221 in backboardd; they were removed without relaxing policy. R372 retains the iOS QoS frontend and qualified driver and has no process crash, but an indexed-draw variant exceeds its unchanged20s translation budget. R373 limits the existing process pipeline-making budget to one call in flight, retaining all shader work and limits. It has no compile-budget failure and reaches the first UI command-buffer callback(31draws/5blits), then64KiB command allocation fills64 BO handles at72,298,496 mapped bytes. R374 ABI1.6 expands BO handles and submit pins to256, with a separately allocated table to satisfy IOKit16KiB object size. Legacy288-byte64-handle submissions and new1056-byte256-handle submissions are independently bounds-checked; GPUVA512MiB and per-BO256MiB limits remain. Actual allocator host regression fills64 before and256 after, rejects overflow, and frees all descriptors. Kernel and userspace build successfully; hardware UI qualification remains pending. Signature and mmap policies remain unchanged. Physical observation remains deferred. The current Mesa and Metal frontend patches plus the E5 entry source are retained in this repository. A separate e5-metal=1 UI candidate publishes NativeE5Accelerator under the original IOAcceleratorES metaclass after loading IOGPUFamily and checking its extent. A read-only bindfs union exposes the owned NativeE5Metal bundle at the standard System extensions path. Four named UI processes receive the existing OS-profile IOKit grant; no UI file-mapping override is enabled. Discovery and original UI rendering are pending board qualification. The E5 Darwin PanVK changes are archived in patches/mesa/e5-panvk-darwin.patch against Mesa efd07bab401b263a9ac114560ecdb9a2934f5d43; clean-index application passes. Explicit unsupported DRM stubs and the audited legacy weak-import fixer are retained under scripts/gpu. Authored API probes and shaders now live in scripts/gpu/vulkan and scripts/gpu/metal. scripts/e5_gpu_api_probe_build.py rebuilds either API probe, validates shaders and signatures, and never stages or boots it. The Vulkan rebuild exactly reproduces R348 executable SHA256 0e799aed35d74e9c0fd8b0426f84d355be56e7d2eec74c4337556a63676dc75c.

Completion criteria

  • MMU/BO lifetime and genuine JM completion/recovery pass, clear and triangle pixels match references across repeated submissions, and GPU-produced images reach the physical scanout.
  • Metal/IOSurface integration and original Settings content are separate UI acceptance milestones.

Current limits

  • The Vulkan/Metal probes, shared-memory cases and native UI do not establish complete API coverage, every app, concurrent CPU/GPU coherence or physical direct scanout.
  • Sustained smoothness, the ABI1.7/notification candidate, runtime GPU power management and memory expansion remain unqualified pending their own evidence.

Next action

  • Settings functional qualification is complete (r390). Reduce latency with one variable per round: either batch/pipeline cache changes that bound recompiles, or submit-path changes that avoid full synchronous waits, each with host regression plus flash evidence.
  • Then prioritize power management and Wi-Fi/Bluetooth; continue structural GPU improvements backed by measurements.

Evidence

Network

planned

Hardware and dependencies

  • Stock Android uses Unisoc sprd_wlan_combo.ko (SC2355 backend) and wcn_bsp.ko; runtime chip ID 0x2355B001 identifies Marlin3 Lite. Parent compatible unisoc,marlin3lite_sdio; WLAN child sprd,sc2355-sdio-wifi.
  • Wi-Fi host: sprd,sdhci-r11 at physical 0x22220000, size 0x1000, 4-bit SDIO; mmc2 negotiates SDR104. DT interrupt SPI 202 (GIC INTID 234), level high. WCN power/reset/wake GPIOs and PMIC rails are shared with BT/FM/GNSS.
  • Actual firmware /odm/firmware/wcnmodem.bin (947120 bytes, SHA256 780c448afbaa614efdb1ffc71ffb6bc05b91d34f752d7994569d5c2dbf1c2a41); Android selects wifi_board_config.xpe.ini (7395 bytes, SHA256 3ff60f8e6756e839cf51531d52965dc43446492d843c749da60f1ad70ac7fde6).

Configuration

  • No Wi-Fi component is selected in board.toml. Ath12kWifi targets different hardware and has been removed from this area; a Unisoc WCN/SC2355 implementation and XNU network interface are required.
  • Vendor modules: sprd_wlan_combo depends on cfg80211, wcn_bsp and sipc-core; wcn_bsp depends on sipc-core and trusty-ipc. These combo-module dependencies do not change this board's observed SDIO transport.
  • Android firmware search path is /vendor/firmware,/odm/firmware. Preserve the XPE board configuration, regulatory data and per-unit factory addresses when preparing local driver inputs; do not substitute AA configuration or generate calibration values.

Procedure

  • In Android, run python3 tools/e5-wcn-inventory.py --output <record.json> --extract dumps/android-wcn-private from e5-xnu; the capture reads modules, bindings, WCN boot messages and vendor blobs without changing radio state.
  • Compare the live /sprd-marlin3 node, SDIO host and module aliases with the extracted module metadata; pin matching SC2355/WCN source and firmware before implementing board hooks.
  • Implement WCN power/reset/clock and SDIO CMD52/CMD53/IRQ handling, then reproduce firmware boot, CMD_SYNC_VERSION, CMD_DOWNLOAD_INI and CMD_GET_INFO before network registration.
  • Add the XNU networking/802.11 adapter, then measure scan, association, key installation, DHCP, bidirectional traffic and repeated stop/start before claiming device support.

Development findings

  • Android confirms SC2355 SDIO with chipid=2355b001 and firmware MARLIN3_20A_RLS2_W24.17.7 / sc2355_marlin3_lite, dated 2024-04-28 15:41:09. The generic combo module also contains SC2332/SIPC and SC2355/PCIe aliases; those are not evidence of this board using either bus.
  • The DTS firmware partition /dev/block/by-name/wcnmodem does not exist on this device. The extracted ODM firmware size matches the 947120-byte successful Android download; use the actual ODM file as the input.
  • Android selects Marlin3 Lite chip index 2 with IPD(1), parses wifi_board_config.xpe.ini and sends CMD_DOWNLOAD_INI. Boot logs also contain command-version mismatches and an UNKNOWN_ERROR response for the third INI section; a future port must handle firmware responses rather than assume every section is accepted.
  • The TSX symlink targets missing /mnt/vendor/productinfo/wcn/tsx_bt_data.txt; Android logs the default 0xffff. /mnt/vendor/wcn is empty at capture. This does not establish an absent RF calibration requirement; board INI, efuse and factory data still need auditing.

Completion criteria

  • Host resources and shared WCN power ownership match the stock board; SDIO functions and IRQ/data transfers work.
  • Firmware reaches its version checkpoint and SC2355 command/event framing, credits, board configuration and regulatory setup work.
  • XNU scans and associates with the chosen security mode, obtains an address, transfers packets in both directions and recovers across repeated stop/start cycles.

Current limits

  • Only Android inventory and blob extraction are verified; no XNU WCN/SC2355 driver or networking acceptance yet.
  • Source equivalence to the extracted dirty vendor revisions, SDIO DMA/cache constraints, command framing, firmware/efuse calibration handling and the target iPadOS 802.11 ABI still need implementation/audit.
  • Factory addresses stay local. Bluetooth, FM, GNSS and the 5G modem are separate follow-up work; shared WCN power handling must preserve their ownership.

Next action

  • Compare the six hashed source files at sprd-linux-devs/linux b93bbc3a36fd3c1acae0e44b93d909f7fe942182 with the extracted vendor module metadata and command ABI. Then implement board power/reset, SDIO function/IRQ transfer and the WCN firmware version checkpoint as the first milestone.

Evidence

Audio

planned

Hardware and dependencies

  • Unisoc audio DSP path unknown; bar-phone speaker and receiver present.

Configuration

  • No XNU audio component or route is configured.

Procedure

  • Read the live Android audio graph and bound codec/DSP drivers; record firmware, clocks and route ownership before implementing playback or capture.

Development findings

  • Audio playback and capture have not been attempted in XNU.

Completion criteria

  • Each declared route carries the selected audio format through playback or capture and handles start and stop correctly.

Current limits

  • No audio work yet.

Next action

  • Dump the vendor audio graph (ADSP, codec, routing) from the live device tree before selecting components.

Evidence

Collect evidence using the next action.

Power

planned

Hardware and dependencies

  • Removable battery (senior bar phone), PMIC over the ADI bus, AON always-on domain.

Configuration

  • No XNU battery or charging component is configured; bring-up retains the existing watchdog recovery path.

Procedure

  • Read Android PMIC, battery and charging bindings and units; compare a first read-only XNU report against the same device state before adding control writes.

Development findings

  • Volume-up and power keys sit behind the PMIC EIC over ADI and were deliberately skipped for the UEFI input path.
  • R65 uses a standard launchd root policy to disable only com.apple.watchdogd while native IOWatchdog integration is unqualified. The original per-job file has no ForceEnableHack and the original loader is exercised offline with explicit API stand-ins. The UMP9620 600-second hardware test/recovery window remains unchanged and Android A returns. This development policy does not provide watchdogd health monitoring or qualify native power management.

Completion criteria

  • Declared power controls and battery readings match observations; each claimed sleep and resume transition restores its dependent peripherals.

Current limits

  • No power management beyond PSCI SYSTEM_OFF and the LK watchdog reset observed so far.

Next action

  • Map PMIC ADI registers and battery gauge nodes from the vendor tree for a first read-only battery report.

Evidence

Collect evidence using the next action.

Image assembly

implemented

Hardware and dependencies

  • kernel.macho from CI, hardware.dtb, ADT with root volume UUID, packed APFS seed ramdisk, trust cache (cdhash list), E5BootMenu, DarwinBoot, UniXNU.efi and startup.nsh on a 56 MiB FAT payload.

Configuration

  • make-e5-payload.py assembles the payload FAT; mk-boot-payload.py wraps BootShim plus SILICIUM_UEFI.fd plus payload into the 62,926,860-byte boot image flashed to boot_b at LBA 8.
  • The trust cache must be rebuilt whenever any seed binary is rebuilt (ad-hoc cdhashes change every build).
  • The ADT root volume UUID and board.toml root_match_volume_uuid must track the APFS seed container UUID.

Procedure

  • Build seed binaries with the iphoneos SDK, ad-hoc codesign, collect cdhashes with codesign -dvvv, build the trust cache, pack the seed, assemble the payload, wrap the boot image and flash.

Development findings

  • R286 selects the E5 input archive and e5-hid boot argument, with a small report/transition workloop. Declared memory ownership, DMA windows, ADT, firmware, trust cache and media layout retain the R285 values; input runtime and physical tests are recorded separately.
  • Payload FAT layout and boot image size are stable across rounds; the packed ramdisk size varies with the seed contents and shifts the staging window without breaking boot.
  • A stale trust cache produces an AMFI unsuitable-CT-policy panic for PID 1; rebuilding the cache fixed it (flash 7 to 8).
  • Flashes #55-#57 preserve kernel/ADT/EFI/configuration file hashes while replacing the trust cache for the actual signed eMMC-root binaries. Root and boot payload are fully read back before reboot; the static-tool entry correction changes ls/cat signatures and requires rebuilding both numeric aliases and the cache. #58 preserves every file except the kernel and matches the earlier kernel-only repacker byte-for-byte. #59 replaces only the ADT memory property, validated over the entire sidecar. Original-system input assembly and independent full-image/window verification are implemented in e5-xnu/tools/e5-system-root.py.
  • The E5TW repair reproduces CI 305 from UniXNU 066c8d8315 plus the pinned arch/SoC/OS series: all five incoming tee blob hashes match, and all 81 patches apply without recount or fuzz. Patches are exported from staged source with git diff. CI 306 compilation and two full-readback board runs subsequently validate this patch and its persisted ring header.
  • Kernel 306 builds and prelinks at device revision 54347d444c71064ba254a8cb9abd787264cd007a. Its paired complete UEFI boots with the same ADT, trust and original OS/SystemOS inputs, and all 62926860 payload bytes match flash readback. A second controlled experiment retains that kernel and firmware, changes only ADT chosen metadata derived from the exact rehashed OS/SystemOS bytes, and is board-verified through mount-phase-2 and the later tzinit EROFS panic.
  • The second CI-306 payload changes only kernel.macho.adt among nine FAT files and preserves the whole firmware prefix. All 62926860 bytes match flash readback. The common profile/board ADT generator exactly reproduces its tested metadata, and the payload is retained as the new staging baseline.
  • The next source build adds e5-data-progress.patch after the existing device series; all 82 pinned patches apply without recount or fuzz. It preserves exit semantics and records tzinit only after P_LEXIT/p_xstat commit. Data size changes to 128 MiB while its UUID is retained. Kernel CI and full payload readback/board qualification are required before accepting this source.
  • CI 315 compiles, links and prelinks source 519a72775fad6b6fd71c1ffbb2b68e41c4838d2e with all 82 pinned patches. Kernel UUID D79DCFE4-737E-3041-A570-388C447577D9 matches board panics in rounds 63/64. R63 changes only kernel.macho from R62; R64/R65 retain the entire same boot payload and vary only independently verified Data fixtures. Guest inode uid/gid fields and containing APFS Fletcher64 checksums are adjusted in a copy for the root policy; strict host read-only mounts verify stored owner 0:0 and reproduce the installed image byte-for-byte. The final boot/Data pair is frozen as e5-xnu baseline-data-r65.
  • R109 appends e5-display-submit-trace.patch after the existing display diagnostics and routes probe output directly to E5TW. Full replay of all 86 resolved patches passes on the pinned XNU/common/SoC/prelink inputs. The candidate payload will retain the r108 firmware and actual ADT, add e5-no-vc in its FAT configuration and require complete boot_b readback before reboot. CI and device results remain pending.
  • R109/CI363 qualifies the submission diagnostic on device. Full payload readback passes; all 101 retained read samples show F=1:557a87bd while 42413 video-console character calls are suppressed. Original selectors 3/4/5 have zero calls. backboardd PID 56 repeatedly calls selector 6 (SwapWait), with at least 11911 calls, zero successes and last status e00002d5 Busy; selector 12 (RequestPowerChange) is called once and returns e00002c7 Unsupported. The native power/readiness contract is the next target, not an inferred empty scene. Fresh Android A returns after 665 s; full 128 MiB Data SHA c1744ab4 and protected OS/SystemOS metadata remain unchanged. No physical panel observation or UI presentation claim.
  • After arm_vm_init establishes the phase-1 mapping, the console tee also calls e5_bootlog_putc, a nonblocking RAM-only producer supplied by the E5 storage archive. Early physical-address console output avoids atomic queue operations. An opt-in worker persists console records to a bounded boot_b journal; RAM pstore remains available. R241 and R242 qualify periodic console writes and retention of the preceding boot slot across warm watchdog resets: 86 and 73 valid records, no reported dropped bytes, CRC failures or stream gaps. Cold power loss and full slot rotation remain unqualified.

Completion criteria

  • Kernel and prelink succeed, payload contains all declared inputs, and the boot image matches the bootloader format and size limit.

Current limits

  • The boot image is built on the host and flashed from Android; no on-device packaging yet.

Next action

  • Automate the seed-to-payload chain into one script with a single UUID hand-off to reduce manual ADT patching.

Evidence

Diagnostics

device-verified

Hardware and dependencies

  • Persistent pstore ramoops zones (dmesg-ramoops-0..5, console-ramoops-0) surviving the watchdog reboot into Android; e5-uplink fd 229 write mirror; E5FB rendered-text mirror; E5KINIT register read-back.

Configuration

  • tools/e5-uefi-cycle.sh collect pulls the pstore zones into dumps/uefi-report-HHMMSS.txt after each flash cycle.
  • The E5UL uplink lines are written from the seed shell through the kernel write hook with the E5UL prefix and reach the persistent ring.

Procedure

  • Flash, wait for the watchdog reboot into Android, then collect and grep the report for the E5 markers.

Development findings

  • dmesg-ramoops-0 is a circular ring holding multiple boots; per-build markers (INIT-START-vN) are required to separate fresh output from residue.
  • Kernel printf stays buffered without timer interrupts; the e5_panic_print tee path is what reliably reaches pstore.
  • Raw wait status matters: #55/#56 print EXIT 0 for a child killed by SIGBUS, but EXITH is 0x0a. Cat startup traces locate the failure before C entry; #57 aligns the C stack while preserving the original argc/argv pointer and all five commands exit 0. Read errors in cat now return exit 67 instead of silently reporting success.
  • The E5TW tee header at unaligned PA 0xFFF81F1B uses volatile byte stores and publishes the next-write position for every character. Its literal magic is 0x57543545, matched by the UEFI reader. Actual writer/reader tests retain the last 4 KiB across multiple ring wraps and short final panic text; ARM64 instruction checks reject unaligned word stores. The EDK2 reader module builds; two board runs now verify raw-ring panic capture and the BootShim E5TW reader branch.
  • CI 306 and the first E5TW board run recover the complete ordered 24568-byte raw console ring at pstore offset 0x1F00, position 16307. Kernel UUID F6A3666A-8868-3E30-BC01-5D1A69866445 matches the artifact. The actual panic is mount[3] exit 66 after missing boot-manifest-hash. Existing hooks triple panic characters, so decoding the full ring is required; a 4 KiB printed tail alone lacks the main panic line.
  • Two independent board sessions recover complete E5TW raw rings at positions 16307 and 19018 with the same expected kernel UUID. The first captures mount[3] exit 66; after the controlled ADT repair the second captures tzinit[17] exit 30. Raw pstore, ordered bytes, normalized text, source hashes and fresh Android boot IDs are retained per run.
  • Round 63 adds persistent in-kernel tzinit wait-status and pid fields, recorded by the accepted process exit thread and echoed in sampled NativeSd t=pid:raw_wait_status rows. ffffffff means unseen; 0 means normal exit zero, while nonzero exit codes and signals retain their native wait encoding. This supplies direct exit evidence even if early console lines wrap. Source patch application is verified; runtime proof is pending.
  • The post-P_LEXIT/p_xstat tzinit hook is now board verified: t=17:0 survives console wrap in rounds 63, 64 and all 248 complete late R65 samples. R63 core output masks the original watchdogd cause; R64 core opening fails with ENOTDIR in 0.086 seconds, preserving the surrounding console without ENOSPC. The original watchdogd user fault PC/assertion text remains unobserved; offline missing-service/open-error branch execution is not mislabeled physical root-cause proof.
  • R109 diagnostic preparation is isolated on fix/display-submit-trace-r109. The original CI362 scanout attribution is corrected in docs/logs/display-content-audit-r108.json; all 86 resolved patches replay with pinned source inputs, recorded in docs/logs/display-submit-r109-patch-check.json. Only kernel diagnostic code and the panel probe change; payload/firmware/Data acceptance remains pending.
  • R109/CI363 qualifies the submission diagnostic on device. Full payload readback passes; all 101 retained read samples show F=1:557a87bd while 42413 video-console character calls are suppressed. Original selectors 3/4/5 have zero calls. backboardd PID 56 repeatedly calls selector 6 (SwapWait), with at least 11911 calls, zero successes and last status e00002d5 Busy; selector 12 (RequestPowerChange) is called once and returns e00002c7 Unsupported. The native power/readiness contract is the next target, not an inferred empty scene. Fresh Android A returns after 665 s; full 128 MiB Data SHA c1744ab4 and protected OS/SystemOS metadata remain unchanged. No physical panel observation or UI presentation claim.
  • After arm_vm_init establishes the phase-1 mapping, the console tee also calls e5_bootlog_putc, a nonblocking RAM-only producer supplied by the E5 storage archive. Early physical-address console output avoids atomic queue operations. An opt-in worker persists console records to a bounded boot_b journal; RAM pstore remains available. R241 and R242 qualify periodic console writes and retention of the preceding boot slot across warm watchdog resets: 86 and 73 valid records, no reported dropped bytes, CRC failures or stream gaps. Cold power loss and full slot rotation remain unqualified.

Completion criteria

  • Each claimed milestone has a reproducible command, a matching source or image revision and an accessible result.

Current limits

  • Reports concatenate several boots; ring rotation can wipe early markers during log floods.

Next action

  • Add a boot-counter prefix to every uplink tag so multi-boot residue is unambiguous.

Evidence

usb-console

device-verified

Hardware and dependencies

  • E5 vendor FDT, e5-linux Qogirn6lite PHY/MUSB/UMP9620 ADI definitions, XNU 10063.141.1 patched baseline.

Configuration

  • Board payload defaults to e5-usb=1; native full-speed polling CDC ACM at physical 0x64a00000, product E5 XNU log, VID:PID 0525:a4a7.
  • Independent 1 MiB kernel log history and 1 ms worker. No EFI callbacks, descriptors or memory pointers are retained after handoff.

Procedure

  • Export the identical shared CDC/PHY implementation with e5-xnu/tools/sync-e5-usb-kernel.py against the prepared preceding patch series.
  • Rebuild the pinned source kernel and prelink, pack a bounded B-slot trial, verify flash readback and collect firmware/native USB plus Android recovery through e5-uefi-cycle.sh.

Development findings

  • The native transport buffers early console bytes before arm_vm_init, then maps its own MMIO and reinitializes controller/PHY. Console producers use a nonblocking lock after VM initialization; USB completion advances the delivery cursor. A startup worker keeps enumeration and idle output alive.
  • Trial 2026110617 qualifies native CDC after firmware handoff on macOS: 76,344 firmware bytes, 34,066 initial kernel bytes and 97,397 kernel bytes after closing/reopening the reader. The kernel stream includes early E5 markers, Darwin 23.6.0, VM startup and BSD root disk4s1. All 25 sampled heartbeats report sent==written and dropped=0.
  • EP0 IN replies must release SETUP RX ownership before filling TX. PIO UNDERRUN is an empty-FIFO NAK, including after a packet ACK; treating it as failure duplicated 64-byte packets. Both transports share the corrected implementation. Normal printf/vprintf is captured at common console_write/unbuffered entries before panel/UART policy.

Completion criteria

  • Native CDC ACM enumerates after ExitBootServices, captures kernel startup and idle heartbeat, survives a reconnect, and the bounded trial returns to Android A.

Current limits

  • Hardware capture used macOS; Linux discovery/reader coverage is host-fixture only. The reconnect test closes/reopens the serial reader, without physical unplugging.
  • A kernel failure before native MMIO mapping cannot deliver its early buffer over USB; panic drain and cold power loss were not separately forced.

Next action

  • Retain the verified USB kernel/payload as a bring-up baseline; separately qualify physical unplug/replug, Linux hardware capture and panic-context delivery.

Evidence

emmc-windows-r257

device-verified

Hardware and dependencies

  • r256 verdict: deterministic console-text/zeros delivered by SD reads (APFS obj_checksum_verify failures) plus a watchdogd crash panic at kernel+970-1000s; user order to move SystemOS+Data to eMMC raw windows inside a shrunk userdata (10 GiB reservation).
  • Shrunk f2fs userdata p70 to 3107840 blocks (GPT end 40103935) with the kanoqwq/openwrt-tools f2fs-online worker; SystemOS container dd'ed to LBA 40103936 (sha b4908cde...) and grown 4 GiB Data container to LBA 46555136 (sha d8b88ba4...), both readback-verified.

Configuration

  • SystemOSWindow.h: E5SystemOS::base 12582912->40103936, E5Data::base 19034112->46555136; sectors, UUIDs and superblock CRC unchanged. Bases now sit past SD capacity so NativeSd::fits() self-refuses and the eMMC host owns both windows.
  • NativeSd.cpp: window refusal no longer fails start(); the host still arms the UMP9620 watchdog (e5-systemos-boot) and publishes the SD card read-only. The SD card must stay inserted: no card means no watchdog arming path.
  • NativeEmmc.cpp: multi-block CMD18 readMulti with whole-range 0xCAFEBABE poison + tail-poison fail (r256 silent-corruption defense), r253 verify-retry windowReadVerified, r254 rw+ big-read trace, per-sector CMD24 writeSingle for Data writes, setDivider/liftClock (capabilities base clock, <=26 MHz legacy ceiling, CMD6 EXT_CSD 183 4-bit + HC1, 4x LBA1 CRC stability checks, fallback to identify clock), E5SystemOSDeviceEmmc/E5DataDeviceEmmc wrapper devices published from start() after superblock probes before registerService.
  • XNU boot-arg panic_on_crash_disabled=1 (existing TUNABLE, kern_exit.c:1695) demotes the watchdogd-exit panic; launchd (initproc) panics remain.

Procedure

  • Commit kext changes, dispatch the kernel builder with this device_ref, fetch kernel.macho via e5-xnu tools/e5-kernel-fetch.sh.
  • Pack the r257 trial from the accepted USB artifacts with --extra-arg panic_on_crash_disabled=1 and the watchdog policy; flash, verify readback, run one bounded cycle with native USB capture, collect journal+pstore.

Development findings

  • Latest R284 evidence supersedes the early r257/r261 bring-up failures: current windows at OS32018432/SystemOS43888768/Data50339968 boot original userland with no retained integrity/panic markers and all 37 protected samples pass. R286 only adds gated HID startup to NativeSd.cpp; the tested window geometry and storage transfers are preserved.
  • R257 brick + recovery; R258 first eMMC-windows boot (journal bytes served to Data reads); R259/R260 liftClock DMA overrun (fixed 858522d); R261 journal-off boot: windows mount, SystemOS union succeeds, one 16 KiB read at window chunk 6170 corrupted with content that exists nowhere on media while Android reads the LBA clean - corruption is in the XNU read path, per-call deterministic, beyond 1 MiB. See the R255..R261 findings entries in the usb-console area for details.

Completion criteria

  • The bounded OS/SystemOS/Data windows mount with correct UUID/geometry, original userland reads and Data writes complete without retained integrity errors, protected windows pass and Android A recovers.

Current limits

  • v4 multi-block SDMA behavior on this controller is unproven until the r257 log shows readMulti success; the poison tail check turns non-advancing transfers into hard read failures instead of silent corruption.
  • Clock lift is capped at the 26 MHz legacy ceiling; HS200/HS400 tuning is out of scope. If liftClock fails, windows still serve at the identify clock (kilobytes/s) and the scan will not finish in any watchdog window.

Next action

  • r262 corruption map: sparse per-1MiB CRC coverage of both windows + dense fault region + wrong-chunk hex dump; identify the wrong-content source and the positional/time distribution, then fix the read path.

Evidence