Files
UniXNU/docs/devices/nuwa.md
2026-11-15 18:44:12 +00:00

196 KiB

Xiaomi 13 Pro (nuwa)

wh1stl3/UniXNU-nuwa

Xiaomi 13 Pro (SM8550), official step 7. 2026-11-15 milestone: the native UFS root volume mounts (native-ufs-root=1, no ramdisk, boot-uuid DADD93F8-...-0F11E64172F7) and launchd reaches user space; userdata's first 8446279680 B were written and read back as da429da6...68d7, and boot_a/vendor_boot_a now carry the payload. The launchd path runs on three opt-in bypasses (root auth skipped, runtime trust cache empty, initproc exec policy ignored); authentication is NOT REACHED. Host artifacts are self-checked and the device observation is parent/user-reported. The interaction target is not met: launchd completion, the display service and reboot-to-same-system are unverified.

Source record: porting.toml, a840aaebb96dd3624c4d06f35f23960664fded64.

This page is generated from the device record.

Area Status
Configuration planned
Firmware and boot planned
CPU and interrupts planned
Memory and DMA planned
Storage and root planned
Display planned
Input planned
GPU planned
Network planned
Audio planned
Power planned
Image assembly planned
Diagnostics planned
cellular planned

Configuration

planned

Components: ArmDTPlatform, ArmCpu, XnuPrelink, DarwinBoot.

Hardware and dependencies

  • SM8550 common repository commit (pinned in feeds.toml).
  • The iOS 16.7.16-20H392 OS profile at UniXNU/profiles/os/ios-16.7.16-20H392, authored locally in UniXNU commit 693f5d9cae8a408dd18b39f3f9fdc4228c934241; the official profiles/os tree ships iPadOS 17.7.11, iPadOS 26.5.2 and iOS/iPadOS 27.0.1 only.
  • The donor KernelCache member hash for [ipsw.sha256]; the profile already carries the IPSW url, chip_id, board_id and the DeviceTree plus StaticTrustCache member hashes read from the donor BuildManifest.

Configuration

  • board.toml schema_version = 2 with [board], [components], [kernel], [device_tree], [firmware], [payload], [payload.adt], [package].
  • feeds.toml pins UniXNU-sm8550-common at a9f45dc60c035084b95463781538380f4c3606c9.
  • dts/nuwa.dts includes soc:dts sm8550.dtsi and carries the measured DRAM span only.
  • board.toml [ipsw] declares product, version, build, device_class and fetch only; url, chip_id, board_id and the member hashes are inherited from the selected profile instead of being duplicated.
  • [payload] tools = true adds the Tools entry to the generated payload boot menu.

Procedure

  • ./unixnu defconfig nuwa_defconfig -D ../UniXNU-nuwa
  • ./unixnu fetch -D ../UniXNU-nuwa
  • ./unixnu status --json -D ../UniXNU-nuwa
  • ./unixnu tree -D ../UniXNU-nuwa

Development findings

  • 2026-11-15 device configuration state that reached launchd: patches/xnu/series mounts 8796-vfs-getvfs-with-vfsops.patch (the APFS root needs _vfs_getvfs_with_vfsops), while the three launchd-path bypass patches stay opt-in outside the series; board.toml [os_overrides] root list now also pulls com.apple.kec.pthread so the donor pthread prelink is bound. No pin, feed or OS profile value changed.
  • 2026-11-13 VFS bridge registered; isolated native object symbol, header and export list plus six actual-source contract cases self-checked. This does not discover a root device; concurrent lifetime and final kernel export remain unverified.
  • Doc-path hygiene: board.toml and porting.toml now reference the live docs (docs/context-brief.md, docs/decisions.md, docs/routing.md); no configuration value changed.
  • The local iOS 16 OS profile lives at UniXNU/profiles/os/ios-16.7.16-20H392 (UniXNU commit 693f5d9cae8a408dd18b39f3f9fdc4228c934241): xnu 8796.141.3 (1b191cb58250d0705d8a51287127505aa4bc0789), the donor IPSW url/product/build/device_class (d22ap, chip 0x8015, board_id 6), the DeviceTree and StaticTrustCache sha256 pins, ADT identity D22AP, kernelcache sha256 and the 31 prelink roots verified against the donor __PRELINK_INFO; board.os_profile and [device].os_profile select it.
  • board.toml sets [firmware] preflight and [bootimg] dtb = 'build/nuwa/sm8550-xiaomi-nuwa.dtb'; the dtb path is relative to the build tree, the base ctx.tree/dtb does not expand a {uefi} template, and an FDF-embedded DTB does not automatically become the appended DTB.
  • Official step 4 is complete: defconfig, fetch, status --json and tree resolve the board against that profile, and the fetch is recorded on 279c990a0c7a238aa0a5818493c6cd6a46528749 (unixnu.lock: xnu 1b191cb5, sm8550 a9f45dc6, edk2 cfb93724, libplatform a00a4cc3, libdispatch 30fe6a7d, libpthread c0e60fb5, AvailabilityVersions 933f9360, corecrypto-2023 11fba65b, IOStorageFamily e45fbe30). Step 5 (hardware description and firmware) is in progress.
  • The official profiles/os tree ships iPadOS 17.7.11, iPadOS 26.5.2 and iOS/iPadOS 27.0.1 only, which is why this profile is authored locally; profile selection is no longer blocked.
  • Kernel patch set (2026-11-08, host-only, self-checked): on xnu-8796.141.3 (1b191cb58250d0705d8a51287127505aa4bc0789) 13 of the 56 arch patches fail plain git apply, so board.toml [kernel] patch_series now mounts only device:patches/xnu/series, a 17-entry minimal set for stage 5: 13 arch patches by arch: reference plus four local 8796 ports (apple-only-cpu-tunables, non-apple-iommu-bootstrap, source-machine-layer, zzx-pmap-sync-64bit-pa). The optional storage, GPU, AES, exec, systemos, cryptex, trust, userland and mockfs patches are not mounted; pgz-aligned-validation and published-freezer-condition do not apply because their target code is absent in 8796. The set applies cumulatively to a clean 8796 clone and unixnu status --json resolves the same order; nothing has been compiled and the four ports plus four high-risk arch patches (native-4k-bootstrap-capacity, native-4k-user-abi, non-apple-cache-monitor, non-apple-pmap-field) have not been reviewed. Later (2026-11-08): nuwa_defconfig adds CONFIG_XNU_NATIVE_AES=n and CONFIG_XNU_SOURCE_PTHREAD=n, board.toml replaces the kernel archive list and points [kernel.prelink] at the 8796 kernel file; the kernel now builds and links, and prelink stops on an ABI gap against the donor kernelcache (see packaging). Update (2026-11-08): [kernel.prelink] roots now exclude apfs and Libm through board.toml and prelink stops at an external PC-relative encoding at 0xfffffff005b8c1cc.
  • board.toml [payload] tools = true adds the Tools entry (id = tools, action = tools) to the generated EFI/UniXNU/unixnu.toml, so the exported firmware log can be read from the boot menu; the entry has not been exercised on hardware.
  • board.toml [firmware.defines] now sets ENABLE_SECUREBOOT = 0 (2026-11-07): SiliciumPkg.dsc.inc:13 passes -D ENABLE_SECUREBOOT=$(ENABLE_SECUREBOOT) unconditionally, and unixnu runs the EDK II build directly instead of Mu-Silicium's build_uefi.py (default 0, --enable-secureboot is opt-in), so an empty macro failed the preprocessing of SecureBootHandler.c:216 (#if ENABLE_SECUREBOOT == 1). Self-checked with the define in place: the firmware build recorded in the firmware area exited 0, and the Verifier subagent later re-ran it independently; nothing has run on hardware.
  • dts/nuwa.dts comment corrected (comment-only): scripts/unixnu/devicetree.py takes exactly one /memory node with one contiguous reg range, so the previous nine-segment TODO was wrong; the 0x8c0000000/0x240000000 geometry and the owned-window contract are unchanged (docs/step5-firmware.md).
  • Repo hygiene: removed the committed __pycache__ copies of the format-1 legacy tooling; no source behaviour changes.
  • Repo hygiene: the committed __pycache__ blobs were untracked and are ignored; scripts/ keep format-1 legacy tooling that still points at deleted absolute paths and is not part of the official flow.
  • The SM8550 shared series targets XNU 10063/12377, so soc:patches/xnu/series is deliberately not selected for the 8796 line.
  • [userdata] is populated in board.toml from the 2026-11-07 read-only GPT capture (docs/device-evidence-2026-11-07.md): first_lba = 3110976, last_lba = 61850618, lba_bytes = 4096, copied from the GPT dump parsed on the host; no GUID, serial, kernel command line content or chip seed is recorded.
  • ADT identity values decided 2026-11-07 (docs/step5-firmware.md, 'ADT identity values'): ADT /chosen/chip-id (u32) and unique-chip-id (u64) are software identity seeds for IOPlatformUUID, not Apple silicon IDs; the generator checks only nonzero and width and never compares them against the SoC. The donor iPhone10,3 DeviceTree (base out/ipsw/20H392/DeviceTree.d22ap.dtb, an LZFSE-compressed ADT) carries chip-id = 0 and unique-chip-id = 0, template placeholders that iBoot fills at run time, so the donor offers no value to copy.
  • chip_id = 4acab8db was drawn at random on the host on 2026-11-07: nonzero, and different from sheng's 0x8142 and from the donor ApChipID 0x8015. It is written in board.toml [payload.adt] and stays fixed once images exist.
  • The unique-chip-id u64 seed is drawn at random on the host as eight hex digits and kept only in the git-ignored unixnu.local.toml of the base UniXNU checkout; no seed value is committed. Every build must reuse the same value, and if it is lost the same device gets a different IOPlatformUUID, so the user must back it up.
  • root_match_volume_uuid is commented out in board.toml for bring-up stages 1-5: fastboot boot uses the in-memory Data seed, the generator omits /chosen/root-matching when the value is empty, and a non-empty placeholder fails UUID parsing. Before stage 6, read the OS APFS volume UUID (not the container UUID and not the Data volume UUID) from the built userdata image and write it to board.toml and to the OS profile [root] os_volume_uuid.
  • Self-checked (2026-11-07, in memory, no device): generating the ADT from the real board.toml, the iOS 16 profile and the local seed returned 12984 bytes; /chosen carries chip-id = 0x4acab8db, an 8-byte nonzero unique-chip-id matching the local seed, dram-base = 0x8c0000000, dram-size = 0x240000000 and darwinboot-memory-size = 0x40000000, has no root-matching (expected at this stage), lists 8 cpu reg entries 0..1792, timebase-frequency = 0x124f800 and psci = smc, and arm-io holds only the intc child because nuwa.dts still describes CPU, memory and intc only (step 6 work). ephemeral-storage is absent because no Data seed was passed, which is a property of this run, not a stage 1-5 conclusion.
  • board.toml [os_overrides."ios-16.7.16-20H392".kernel.prelink].roots drops 4 of the profile's 31 roots for stage 5 (apfs, Libm, IOAVFamily, IOMobileGraphicsFamily); dropping IOAVFamily/IOMobileGraphicsFamily removes the transitive Libm dependency, and prelink then completes (host build, exit 0, 55,357,440-byte kernel.macho); host build only, self-checked.
  • ADT identity TODO filled from the 20H392 (16.7.16) donor DeviceTree (Firmware/all_flash/DeviceTree.d22ap.im4p -> decoded 'bvx2' ADT at out/ipsw/20H392/dt.dtb): board.toml [os_overrides."ios-16.7.16-20H392".adt.product] now carries product-id = fd079b298144a7deb4d7c87fef3eb474d8f0e7eb (20-byte MGProductType, dt.dtb offset 0x29544), artwork-device-subtype = 2436 (u32, dt.dtb offset 0x29e48) and corrects compatible-device-fallback from the profile's 'iPhone10,1' to 'iPhone9,2' (dt.dtb offset 0x29698); device-class (d22ap) identity, host-only, self-checked.
  • early_security_init and text_fixups stay unset for stages 1-5 (recorded as skipped, not required): the 8796 early security initializer is a reversing task on the stripped AMFI kext (not a field lookup) and on a non-Apple SoC without an Apple PMGR running it risks the 'AMFI: No PMGR?' panic (same SoC family as nx769j/caihong, UniXNU docs/devices/nx769j.md); text_fixups are dyld-cache byte fixes only derivable when a boot reaches one (nx769j: 'there is nothing to derive before that'). AMFI still starts through its ordinary start routine without them.
  • board.toml [payload] size_mib changed 192 -> 64: the prelinked kernel (~53 MiB) plus trust cache and ADT needs a FAT volume that fits the 96 MiB vendor_boot partition (192 MiB would not fit); 64 MiB keeps the volume-check headroom.
  • board.toml [payload] now pins trust_cache_sha256 = ["367e051548f9f0407fa9fe9d40007fa4e7c5ff27c45f5a7a30cedf344e24d98e"] (a device-level array replacing the 20H392 profile's raw-hash pin 63e01c03… as a whole): the profile pins the raw StaticTrustCache IM4P hash, but DarwinBoot's ValidTrustCaches (DarwinBoot.c:371-397) only accepts the bundle format, which is why the on-device xnu entry returned EFI_COMPROMISED_DATA; the same donor module re-bundled with scripts/bundle-ipsw-trust.py is pinned instead. Host-only (self-checked).
  • Entry probe for stage-4 diagnosis: board.toml [payload] panic_screen = true makes payload.build patch a copy of kernel.macho (scripts/patch-xnu-entry-probe.py --panic-screen --no-reset) - stage 1 blue bar at the MMU-off kernel entry, stage 2 magenta bar just before SCTLR_EL1, plus the _panic/_panic_trap_to_debugger redirect; no reset, so the panel stays lit and the device is recovered by holding power (PSCI SYSTEM_RESET would cold-boot Android). Host build only (self-checked).
  • [payload] entry probe variant for stage 3 (2026-11-09, host build, self-checked): board.toml now sets entry_probe_reset_delay = 6 and panic_screen = false. Enabling panic_screen together with the reset delay makes the panic screen code (0xa2f bytes for nuwa's kernel) overflow the zero padding after _start by 0x5f (95) bytes past PANIC_CAVE_END, which scripts/patch-xnu-entry-probe.py refuses (same limitation as the alioth device record). This variant keeps the stage 1 blue bar, stage 2 magenta bar and the 6 s stage 3 PSCI SYSTEM_RESET; the panic path is not redirected.
  • Entry probe stage 3 verified on device and switched to stage 4 bisect (2026-11-09): the stage-3 payload (entry_probe_reset_delay = 6, panic_screen = false) auto-rebooted ~6 s after the bars on hardware, proving the CPU passed the MMU enable (msr sctlr_el1) and kept executing. The project adaptation has not flashed userdata; before that, userdata had already been overwritten with Apple data by other operations, so Android cannot boot (user-reported, not independently verified). After reset, the device gets stuck at the Android boot animation and consumes retries for the active slot; slot A TWRP is unaffected. board.toml [payload] now adds entry_probe_reset_after = "_ml_parse_cpu_topology" (entry_probe_reset_in defaults to _arm_init, occurrence 1): stage 4 replaces stage 3 (payload.py:169-177 passes --reset-after; patch-xnu-entry-probe.py:389-396 emits exactly one site via elif, so the two never coexist and there is no masking). The 6 s timed PSCI SYSTEM_RESET now fires after the single _arm_init call to _ml_parse_cpu_topology (bl at 0xfffffff007998914, occurrence 1) instead of right after the MMU enable; a reset then means the call returned, no reset means it did not return (or was never reached).
  • Stage-4 next probe (2026-11-09, host-only candidate, self-checked): user-reported hardware reset after about 6 s with revision 07e176761020d3d7f01fb086d9a89af3e3531eb9 proves _ml_parse_cpu_topology returned. board.toml now selects _arm_vm_init in _arm_init, occurrence 1, delay 6, panic_screen=false; reset means arm_vm_init returned, no reset is inconclusive (not reached, did not return, or probe/reset failure). Cave 4 at 0xfffffff0073ee620 is inside kernel __TEXT_EXEC; patch-xnu-entry-probe.py:97-118 uses PC-relative BL/literal load, a 115200000-tick constant, register-only CNTPCT_EL0/PSCI SMC. xnu-8796.141.3 osfmk/arm64/arm_vm_init.c:1998 obtains the KVA segment, :1413 maps the full __TEXT_EXEC ROX, :2097 calls arm_vm_prot_init, and :2144-2156 switches tables and invalidates the bootstrap identity map. Identity-map removal alone therefore does not invalidate the cave; actual post-return execution remains untested. Existing payload/vendor_boot image and JSON .v5 backups were retained without overwrite. Both build commands in the packaging procedure exited 0; the payload is 64 MiB <96 MiB, while stock-padded vendor_boot is exactly 96 MiB, NOT strictly smaller. No device action or push.
  • 2026-11-09, user-reported hardware evidence relayed by the parent session: the six-second _arm_vm_init post-return probe at device revision 2d55e2a8d223f558f01c948db45db3c71f747c96 reset on hardware. This establishes that callee returned; it does not establish that subsequent code must hang. The implementer ran no device command this round. Next diagnostic candidate (2026-11-09, host-only, self-checked; Reviewer PASS WITH CONDITIONS supplied by parent): board.toml [payload] panic_screen=false, entry_probe_reset_delay=6, entry_probe_reset_after=_PE_create_console, entry_probe_reset_in=_arm_init, entry_probe_reset_occurrence=1, entry_probe_reset_before=false. The actual BL site is 0xfffffff007998b84; cave 0xfffffff0073ee620 lies in __TEXT_EXEC RX. _PE_create_console tail-calls initialize_screen, which can return. The six-second delay starts after the callee returns, not at boot or at the magenta bar. A probe reset proves only this callee returned, not that the following PE_init_printf ran; no reset is inconclusive (not reached, did not return, or probe/reset failure). The BCM2837 fixed panic-buffer access at 0x3c0fc000 in pexpert/arm/pe_init.c remains a risk and is not changed. No VM, memory-window, console implementation or base-source change, no device operation and no push; all area statuses remain planned. This finding supersedes historical claims that no reset necessarily means no return. New candidate not hardware-tested; independent verification pending.

Completion criteria

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

Current limits

  • The donor KernelCache member hash (kernelcache.release.iphone10b) is not pinned in [ipsw.sha256] yet; the OS and Cryptex1,SystemOS member hashes are taken at the userdata stage.
  • No board DTS register set for storage, display, input, GPU, network, audio and power yet, so those components stay unselected.
  • ADT identity: chip_id is decided and written (4acab8db); root_match_volume_uuid stays unset until the userdata image is built before stage 6; the unique-chip-id seed lives only in the base checkout's git-ignored unixnu.local.toml and needs a user backup.

Next action

  • Official step 5: finish the uefi-nuwa firmware integration (UniXNU, UniXNUDxe, DarwinBoot, RamDisk, DiskIo and FAT modules plus the Advanced Logger export) and reconcile the board DTS with the firmware DTB before selecting step 6 drivers.

Evidence

Firmware and boot

planned

Hardware and dependencies

  • Identify the firmware and boot hardware, resources and existing component interfaces for this board.

Configuration

  • Record the board.toml fields, DTS properties and defconfig values used for firmware and boot.

Procedure

  • Follow the firmware and boot stage in the porting guide and record the actual commands.

Development findings

  • 2026-11-15 boot chain is device-resident: boot_a carries the r4 Mu-Silicium loader plus DarwinBoot (1495040 B, 692cfcca51af81230fecbba15425cf598182fbe53abb362b6046dec0f24dc2e5) and vendor_boot_a the launchd-path package (r9, 37b665da4629f69ad97fca1335c91f99f81780c896ebceac4dd419264de8667f), both written under this round's authorisation with the slot active, init_boot_a, vbmeta, misc and userdata untouched; stock images and per-round rollback steps are in out/nuwa-boot-payload-write-20261115-01/. The writes have bootloader write acknowledgement only (no fastboot fetch, no adb in fastboot mode).
  • Doc-path hygiene: board.toml and porting.toml now reference the live docs (docs/context-brief.md, docs/decisions.md, docs/routing.md); no configuration value changed.
  • board.toml [os_overrides."ios-16.7.16-20H392".kernel.prelink] also replaces roots (27 of the profile's 31; apfs, Libm, IOAVFamily and IOMobileGraphicsFamily dropped for stage 5; IOAVFamily/IOMobileGraphicsFamily return with the GPU/media step, and dropping them removes the transitive Libm dependency whose literal pool still trips the prelink planner); host build only, see packaging.
  • board.toml now also carries [os_overrides."ios-16.7.16-20H392".kernel.prelink] kernel = the linked mach.development.bcm2837.unstripped (the profile names the 10063 .unstripped.noctf file, which the 8796 makedefs never produce), and a replaced [kernel] archives list; see the configuration and packaging areas. Host build only.
  • board.toml [os_overrides."ios-16.7.16-20H392".root] os_volume_uuid is empty for stages 1-5: the OS profile placeholder REQUIRED_ROOT_VOLUME_UUID stops the kernel bootstrap pass, the owner boot manifest is off for nuwa, and the real OS APFS volume UUID is read from the built userdata image before stage 6; host build only.
  • board.toml [kernel] patch_series now mounts only device:patches/xnu/series (the minimal 8796 patch set, see the packaging area); this is a build-configuration change and was not run on hardware.
  • Reviewed firmware facts (2026-11-07, docs/step5-firmware.md): the firmware DTB carries no splash_region node, while the live machine reports 0xb8000000 with size 0x2b00000 and MemoryMapLib already describes a Display Reserved region. Whether ABL consumes that reservation is unproven, so nothing is injected into the DTB by this review; the firmware repository carries a separate DSC/FDF integration (uefi-nuwa commit becd65b921553957899cc069d6ac6c1278dedd08).
  • Vendor inputs: the nuwa DXE set comes from the public Project-Silicium/Device-Binaries repository, path nuwa/, commit 4648d229f9bedf0d17ba6fe2bebd3fc6fb01422d (git tree d8512b51f264cb4dc2114a788066a80fb77cf8b8, 333 files, 8.5 MB), checked out sparsely into uefi-nuwa/Binaries and kept out of the firmware repository history. A local per-file SHA-256 manifest hashes to 6cff995a23b458bcbee0518dbe9b9d74eeea388e410b17413c913a361e702cab. With this tree present every Binaries/nuwa reference in nuwa.dsc/nuwa.fdf resolves (static path check only); the remaining unresolved references belong to the Mu and Mu_Basecore submodules, whose contents are staged by hand for the build (see the build-input finding below). The repository declares no licence, and the upstream origin of these binaries is unverified.
  • bootimg.dtb wiring stays deferred to a separate report. The ADT identity is decided: chip_id = 4acab8db is written in board.toml (see the configuration area), and root_match_volume_uuid stays commented out until the userdata image exists before stage 6.
  • Step 5 firmware integration is recorded as an independently re-run build (2026-11-07): after an earlier self-checked run (the Verifier subagent's first attempts returned 'Country, region, or territory not supported'), the Verifier subagent re-ran the build against device repository b1064eb8 and uefi-nuwa e494c01b with PASS. uefi-nuwa commit becd65b921553957899cc069d6ac6c1278dedd08 holds the nuwaPkg DSC/FDF integration (UniXNU, UniXNUDxe, UniXNURamDiskDxe, DarwinBoot and AdvancedLogDxe with DiskIo and FAT). From the base checkout (cwd /mnt/d/Code/PortingiOS/UniXNU) `./unixnu build firmware -D ../nuwa/UniXNU-nuwa --uefi ../nuwa/uefi-nuwa --edk2 /home/wh1stl3/nuwa-fw/mu_basecore` exited 0 and produced uefi-nuwa/Build/nuwaPkg/DEBUG_CLANGPDB/FV/SILICIUM_UEFI.fd, 3145728 bytes = 0x300000, matching [firmware] uefi_fd.size in nuwa.toml; FVMAIN.inf lists UniXNU, UniXNUDxe, UniXNURamDiskDxe, DarwinBoot and AdvancedLogDxe. AdvancedLogDxe's producer is now wired in uefi-nuwa commit e494c01b2f679439fdaeaf2084b693c7220543f5: DXE_CORE uses AdvLoggerPkg's DxeCore AdvancedLoggerLib (its constructor installs gAdvancedLoggerProtocolGuid), DXE_DRIVER/UEFI_DRIVER/UEFI_APPLICATION use the Dxe instance and DXE_RUNTIME_DRIVER the Runtime instance, the real AdvancedLoggerHdwPortLib runs over SerialPortLib (DEBUG keeps FrameBufferSerialPortLib so the screen log remains and is also written to the buffer), DEBUG sets DebugLib = BaseDebugLibAdvancedLogger while RELEASE is unchanged, SEC keeps BaseDebugLibSerialPort because Sec.c calls BaseDebugLibSerialPortConstructor directly (the first link attempt failed before that correction), and PcdAdvancedLoggerBase = 0 with the DXE_CORE library allocating a 4 MiB buffer via AllocateReservedPages instead of a fixed physical address. The independent rebuild passed and produced a 3145728-byte SILICIUM_UEFI.fd; DxeMain links DxeCoreAdvancedLoggerLib, AdvancedLogDxe and UniXNUDxe link DxeAdvancedLoggerLib, Guid.xref contains gAdvancedLoggerProtocolGuid, and FVMAIN.inf still lists the five modules. Runtime dispatch and the Tools -> Advanced Logger page remain unverified (a device run is needed), the assumption that the 4 MiB reserved-page allocation lands in the DXE heap is unconfirmed, the boot-chain review concluded APPROVE WITH CONDITIONS on 2026-11-07 scoped to the first fastboot boot (deep not used; no candidate met the upgrade threshold in `../../docs/routing.md`), and the memory ownership gate is still open.
  • Firmware build inputs (2026-11-07): Mu_Basecore comes from the GitHub archive cfb93724a06ff86ff5ba971413368a7809ee42a5 extracted under WSL ext4 at ~/nuwa-fw/mu_basecore; the archive carries no submodule contents, so five submodule directories were copied from the already-checked-out base out/feeds tree (MdePkg/Library/BaseFdtLib/libfdt, MdePkg/Library/MipiSysTLib/mipisyst, BaseTools/Source/C/BrotliCompress/brotli, MdeModulePkg/Library/BrotliCustomDecompressLib/brotli, SecurityPkg/DeviceSecurity/SpdmLib/libspdm). openssl (archive 67b5686b4419b4cb8caa502711c41815f5279751) sits at the uefi-nuwa gitlink path; Common/Mu b8d46c71, Common/Mu_OEM_Sample b70b4b7b and Silicium-ACPI 2d465ad7 come from the fixed pins; Binaries are listed in proprietary-files.txt. Inside WSL the link needs MTU 1400 and the proxy is passed through environment variables to the Windows host gateway port 7890 with allow-lan enabled. These staging steps are manual and not scripted yet; the command sequence is in docs/step5-firmware.md.
  • [firmware] patches wire Resources/MuPatches/Gic-Spurious.patch from the Mu-Silicium checkout (sheng uses the same one); during the 2026-11-07 build (built; independently re-run by the Verifier subagent on b1064eb8/e494c01b) it applied cleanly to the shared edk2 tree and rolled back cleanly.
  • Format-2 migration: [firmware] points at the device Mu-Silicium checkout and proprietary-files.txt lists the required inputs; the vendor input set is recorded and hashed in proprietary-files.txt.
  • board.toml wires [firmware] preflight and [bootimg] dtb = 'build/nuwa/sm8550-xiaomi-nuwa.dtb' (the dtb path is relative to the build tree; the base ctx.tree/dtb does not support a &#123;uefi&#125; template, and an FDF-embedded DTB does not automatically become the appended DTB). scripts/prepare-nuwa-firmware.py only copies the generated DTB and does not edit any node.
  • Self-checked (2026-11-07, not independently verified: the Verifier subagent returned 'Country, region, or territory not supported' on both attempts): in the base checkout, ./unixnu defconfig nuwa_defconfig and ./unixnu tree -D ../nuwa/UniXNU-nuwa both exited 0 with the tree at UniXNU/out/nuwa/ios-16.7.16-20H392/nuwa/tree; in that tree, UNIXNU_UEFI=.../uefi-nuwa python3 scripts/prepare-nuwa-firmware.py exited 0 and cmp showed tree/build/nuwa/sm8550-xiaomi-nuwa.dtb byte-identical to the firmware repository FdtBlob/sm8550-xiaomi-nuwa.dtb; dtc -I dts -O dtb -i UniXNU-sm8550-common/dts nuwa.dts exited 0; and unixnu.devicetree.generate returned 12552 bytes containing the dram-base, dram-size, timebase-frequency, psci, cpu0 and cpu7 substrings (substring presence only; per-value CPU reg and GIC stride were not checked, so they stay unverified). This is a hardware-side resource-encoding check, not a complete ADT: chip_id and root_match_volume_uuid were still placeholders at that run (the later chip_id decision is recorded in the configuration area). No host-class evidence entry is added because it would require a command plus a full 40-character revision, and the device repository changes are uncommitted, so no revision is supplied rather than invented.
  • MemoryMapLib reconciliation (static, 2026-11-07, docs/step5-firmware.md section 'MemoryMapLib reconciliation (static)'; static analysis, not verified on hardware): the static table has 41 entries, 27 inside DDR, while the live FDT has 35 fixed-reg reserved regions, 6 of them no-map regions the table does not describe. If RamManagerDxe pads uncovered System RAM gaps as Conventional, 16 fragments total about 10.80 GiB and a 1 GiB window first-fits in the 0x8c0000000 fragment inside the ADT dram-base. Do not shrink owned_memory_size: below about 925 MiB the window falls outside the ADT range, and any memory-number change needs user confirmation. The RAM Partition content is unknown.
  • Reviewer verdict (2026-11-07, read-only static, docs/step5-firmware.md section 'First fastboot boot: Reviewer verdict and conditions'): window-allocation risk is Medium, APPROVE WITH CONDITIONS for the first fastboot boot only, and Deep Reviewer was not used because no candidate met the upgrade threshold in `../../docs/routing.md`. The failure path writes no partition and prints NO WINDOW/FREE, the success path prints pa/va/entry/args/owned, and XNU's physical range comes from boot_args. Still open: the RamPartition protocol producer (RamManagerDxe is not in apriori, Depex = TRUE, runs at nuwa.fdf:87), SEC HOB write confirmation, a DEBUG build, fastboot boot only with user authorisation, the capture plan, and the FD/ramoops overlap confirmation.
  • Collect the first firmware and boot observation, its cause when established, and the resulting implementation decision.
  • board.toml [payload] size_mib is now 64 (was 192) so the payload FAT volume, which becomes the vendor_boot vendor ramdisk, fits the 96 MiB partition; the ADT identity values (product-id, artwork-device-subtype, compatible-device-fallback) are now carried as [os_overrides."ios-16.7.16-20H392".adt.product] so the generated ADT sidecar matches the 20H392 donor DeviceTree; host-only, no device write.
  • Next diagnostic candidate (2026-11-09, host-only, self-checked; Reviewer PASS WITH CONDITIONS supplied by parent): board.toml [payload] panic_screen=false, entry_probe_reset_delay=6, entry_probe_reset_after=_PE_create_console, entry_probe_reset_in=_arm_init, entry_probe_reset_occurrence=1, entry_probe_reset_before=false. The actual BL site is 0xfffffff007998b84; cave 0xfffffff0073ee620 lies in __TEXT_EXEC RX. _PE_create_console tail-calls initialize_screen, which can return. The six-second delay starts after the callee returns, not at boot or at the magenta bar. A probe reset proves only this callee returned, not that the following PE_init_printf ran; no reset is inconclusive (not reached, did not return, or probe/reset failure). The BCM2837 fixed panic-buffer access at 0x3c0fc000 in pexpert/arm/pe_init.c remains a risk and is not changed. No VM, memory-window, console implementation or base-source change, no device operation and no push; all area statuses remain planned. This finding supersedes historical claims that no reset necessarily means no return. New candidate not hardware-tested; independent verification pending.

Completion criteria

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

Current limits

  • The firmware build was independently re-run by the Verifier subagent on 2026-11-07 against device repository b1064eb8 and uefi-nuwa e494c01b: PASS (exit 0, 0x300000-byte FV).
  • First hardware runs (2026-11-07, three fastboot boot runs, no partition written, read from phone-camera video, see docs/step5-firmware.md): the firmware starts, prints DXE logs and, since uefi-nuwa d7c97a7 (DXE Heap 0xBAB00000+0x0D1C0000), shows the UniXNU menu; Tools -> Advanced Logger exports a log as 24 OneSend QR frames that decode on the host (179 lines); RamManagerDxe maps the high 9 GiB and UniXNURamDiskDxe publishes an initrd block device. Stage 1 and the menu part of stage 2 are met. Stage 3 half met (2026-11-08): with a placeholder payload written to the unused slot's vendor_boot_b, the menu reported config EFI/UniXNU/unixnu.toml and listed the xnu entry (photo; see docs/step5-firmware.md); payload hashes, the ADT sidecar, DarwinBoot and the memory window are not yet exercised and need a real kernel.
  • The Advanced Logger producer is wired (uefi-nuwa e494c01b2f679439fdaeaf2084b693c7220543f5) and the image builds, but runtime dispatch and the Tools -> Advanced Logger page are unverified, the 4 MiB reserved-page allocation's placement in the DXE heap is an assumption, and the boot-chain review concluded APPROVE WITH CONDITIONS on 2026-11-07, scoped to the first fastboot boot only (deep not used; no candidate met the upgrade threshold in `../../docs/routing.md`).
  • The memory ownership gate is open (owner of 0xbe000000-0xd4c00000, tail 0xafffff000-0xb00000000, internal reservations), so the XNU handoff stays withheld.
  • The DTB has no splash_region and ABL consumption of the MemoryMapLib Display Reserved entry is unproven, so the splash reservation stays undecided; [bootimg] dtb remains unset, chip_id is decided (4acab8db) and root_match_volume_uuid waits for the userdata image.
  • Window allocation is Medium risk, not High (static analysis, not verified on hardware): the RAM-partition protocol producer is unconfirmed, XNU's use of the ADT dram-base/dram-size is unconfirmed, and the first boot will show NO WINDOW/FREE lines without writing a partition if it fails. 6 live no-map regions are not in MemoryMapLib, and do not shrink owned_memory_size (below about 925 MiB a first-fit window falls outside the ADT range; memory-number changes need user confirmation).

Next action

  • Boot the image on hardware to verify Tools -> Advanced Logger and runtime dispatch; the boot-chain review concluded APPROVE WITH CONDITIONS on 2026-11-07, scoped to the first fastboot boot only, so it is not a release to boot (see docs/step5-firmware.md).
  • Resolve the RAM Partition protocol producer (gEfiRamPartitionProtocolGuid) by binary reference-point analysis and the first DEBUG boot log; decide with the user whether to add the 6 live no-map regions to MemoryMapLib, and do not shrink owned_memory_size.
  • First boot must use a DEBUG build and fastboot boot only (no partition write; user authorisation in that round), and SEC writing the entry parameters into the HOB (UniXNURamDiskDxe.c:215-233) must be confirmed first.
  • The Verifier re-run passed on 2026-11-07 against device repository b1064eb8 and uefi-nuwa e494c01b; repeat it only if the firmware changes, and keep the recorded command and revisions with the result.
  • Collect memory ownership evidence before the XNU handoff: the active owner of 0xbe000000-0xd4c00000, the 0xafffff000-0xb00000000 tail and the internal reservations.
  • Wire bootimg.dtb and the ADT identity from the separate report, and inject a splash reservation only once ABL consumption is proven.

Evidence

CPU and interrupts

planned

Hardware and dependencies

  • Identify the cpu and interrupts hardware, resources and existing component interfaces for this board.

Configuration

  • Record the board.toml fields, DTS properties and defconfig values used for cpu and interrupts.

Procedure

  • Follow the cpu and interrupts stage in the porting guide and record the actual commands.

Development findings

  • Reviewed CPU facts (2026-11-07, docs/step5-firmware.md): the live MPIDR Affinity1 values 0x000..0x700 for the eight cores and the 3x A510, 2x A715, 2x A710 and 1x X3 cluster mix agree with the board tree, and [payload.adt] cpu_regs = 0,256,512,768,1024,1280,1536,1792 already encodes that numbering.
  • Reviewed GIC facts: GICD 0x17100000 + 0x10000 and GICR 0x17180000 + 0x200000 with an 0x40000 stride match the live machine; the 3-cell versus 4-cell interrupt encoding difference is legal because ArmCpu mapGic does not read interrupt-cells, so addresses and encoding stay unchanged.
  • Reviewed timer facts: the live PPI set 13,14,11,12 differs from the DTB's 13,14,11,10 only in the fourth entry, which is an EL2 interrupt. The current path is EL1 only and CNTV_CTL (INTID 27, PPI 11) does not consume that entry, so the IRQ table is unchanged; the 19.2 MHz reference comes from a live property, not a CNTFRQ read-back.
  • Live CPU re-check, 2026-11-07 (docs/device-evidence-2026-11-07.md): MIDR is 0x411fd461 on cpu0-2 (A510), 0x411fd4d0 on cpu3-4 (A715), 0x412fd470 on cpu5-6 (A710) and 0x411fd4e0 on cpu7 (X3); soc_id 519, machine KALAMA, family Snapdragon, revision 2.0; the live FDT has cpu@0..cpu@700 (eight nodes), psci method smc and compatible arm,psci-1.0.
  • Interrupt and timer re-check: GICD 0x17100000 + 0x10000 and GICR 0x17180000 + 0x200000 with redistributor-stride 0x40000 and #interrupt-cells = 3; the timer node lists PPI 13,14,11,12 with clock-frequency 0x124f800 (19.2 MHz). At Android runtime arch_timer appears as GICv3 virq 11 (virtual timer PPI 11, matching ArmCpu INTID 27); the firmware DTB fourth entry is 10 against the live 12 and the current EL1-only path does not consume it.
  • CNTFRQ was not read back: the boot-time dmesg record had already wrapped in the ring buffer when the capture ran, so cntfrq_readback stays unverified and the 19.2 MHz value remains a live DTB property, not a measured counter frequency.
  • Collect the first cpu and interrupts observation, its cause when established, and the resulting implementation decision.

Completion criteria

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

Current limits

  • The reviewed differences are inert only on the EL1-only path; nothing in this area has run on hardware, so the CPU and interrupt records remain source-level.
  • Hardware inputs and implementation are pending for nuwa.

Next action

  • Keep the DTB PPI set and [payload.adt] cpu_regs as recorded, and validate the EL1-only timer path on hardware once a firmware build exists.

Evidence

Memory and DMA

planned

Hardware and dependencies

  • Identify the memory and dma hardware, resources and existing component interfaces for this board.

Configuration

  • Record the board.toml fields, DTS properties and defconfig values used for memory and dma.

Procedure

  • Follow the memory and dma stage in the porting guide and record the actual commands.

Development findings

  • Reviewed memory facts (2026-11-07, docs/step5-firmware.md): the board DTS geometry 0x8c0000000 + 0x240000000 (9 GiB) is supported by the live high-end pair of adjacent ranges, but the last 4 KiB (0xafffff000-0xb00000000) and the internal/dynamic reservations inside the range are still open.
  • Reviewed memory facts: the current FD window 0xa7100000-0xa7400000 lies inside the historical ramoops 0xa7000000-0xa7400000, and the DXE heap 0xbae00000-0xbf300000 intersects the historical iomem reserved range 0xbe000000-0xd4c00000 whose active owner is unknown. board.toml address, size and verified values stay unchanged; neither permanent Linux reserved ownership nor a full SystemRAM release is inferred.
  • The delivered conventional window comes from RAM Partition -> HOB/EFI map -> DarwinBoot conventional allocation, not from an automatic ADT subtraction of reserved ranges; the XNU handoff stays withheld until the reservation owner question closes.
  • scripts/unixnu/devicetree.py accepts exactly one /memory node with one contiguous reg range, so the nine-segment TODO in dts/nuwa.dts was corrected as a comment-only change; the geometry and the owned-window contract are unchanged.
  • Live FDT re-check, 2026-11-07 (docs/device-evidence-2026-11-07.md): ddr_size reads 0xc (12 GiB) and /memory carries nine reg ranges: 0x80000000+0xe00000, 0x811d0000+0x56e30000, 0xd8140000+0x20000, 0xd8800000+0x0, 0xe7cb0000+0x18350000, 0x880000000+0x39d00000, 0x8c0000000+0xc0000000, 0x980000000+0x180000000 and 0x8b9f00000+0x1e00000. The high pair 0x8c0000000+0xc0000000 and 0x980000000+0x180000000 is adjacent and spans 0x8c0000000..0xb00000000 (9 GiB), matching the board DTS geometry 0x8c0000000+0x240000000.
  • reserved-memory in the live FDT (not an exhaustive list): ramoops@0xa7000000 reg 0xa7000000+0x400000 with no no-map, so it is a real node rather than a placeholder; splash_region 0xb8000000+0x2b00000 (cont_splash_region); xbl_dt_log@81a00000 0x81a00000+0x40000 with no-map; smem 0x81d00000+0x200000. The FD window 0xa7100000-0xa7400000 still falls inside the ramoops range: that is recorded as an observation only, with no conclusion about who owns the range today, no board.toml or DTS change, and neither a reservation nor a release implied.
  • Still missing for the memory gate: evidence for the current firmware version's RAM Partition -> HOB -> EFI memory map -> DarwinBoot window, which needs a UEFI log taken after fastboot boot-ing our own firmware.
  • Static code reading by the main agent, not verified on hardware (2026-11-07, docs/step5-firmware.md): DarwinBoot's window check (UniXNU/drivers/DarwinBoot/DarwinBootPkg/Driver/WindowSize.h:6-8) requires an address in 0x10000000..0x200000000 with 2 MiB alignment and accepts owned_memory_size = 0x40000000; the WINDOW_SIZE default of 256 MiB at DarwinBoot.c:20 is only the parse upper bound; AllocateWindow (DarwinBoot.c:495-546) searches only contiguous EfiConventionalMemory and, on failure, takes the NO WINDOW branch (DarwinBoot.c:531-544) that prints the NO WINDOW/FREE lines on screen; DarwinBoot.c:745-748 sets boot_args PhysicalBase to the window address and MemorySize to WindowSize, so XNU's physical memory range comes from boot_args rather than from the ADT.
  • Static code reading by the main agent, not verified on hardware: the nuwa MemoryMapLib static table (uefi-nuwa Platforms/Xiaomi/nuwaPkg/Library/MemoryMapLib/MemoryMapLib.c:9-35) has one Conventional entry that also produces a HOB, the DXE heap 0xBAE00000+0x4500000 (about 69 MiB), and no high DDR above 0x8c0000000. RamManagerDxe maps RAM-range gaps not covered by the existing EFI map as EfiConventionalMemory with no address or size filter (Silicon/Silicium/SiliciumPkg/Drivers/RamManagerDxe/RamManager.c:48-113), ManageRam cleans up directly when the RAM range lookup fails (:187-215), its depex is TRUE (RamManagerDxe.inf:26-27), and QcomPkg.dsc.inc:30 overrides RamManagerLib with QcomPkg/Library/RamManagerLib, whose RamManager.c:57 returns an error when LocateProtocol(gEfiRamPartitionProtocolGuid) fails. The protocol bytes appear in the nuwa vendor Binaries/nuwa .efi files only once each in EnvDxe.efi and QcomBds.efi, neither in a depex, and the producer is unconfirmed; if the protocol is not installed when RamManagerDxe runs, no high memory is added and DarwinBoot reaches NO WINDOW. This is a silent failure point and is unverified.
  • Static code reading by the main agent, not verified on hardware: no consumer of dram-base/dram-size exists in UniXNU/drivers or the SoC repository (documentation only), and whether the XNU patch side consumes them is unconfirmed; official docs/zh/devices/luming-infinityex-rsxq.md:174 requires comparing the firmware memory map, the DTS memory node, the ADT chosen dram-base/size and the DarwinBoot owned window. Inference (unverified): with a 1 GiB window the low Conventional fragments are likely too small, so a first-fit search would land near 0x8c0000000, and whether XNU accepts a window outside the ADT dram range is unknown. Public reference, not nuwa evidence: official docs/zh/devices/alioth.md:193/200/226 records a Mu-Silicium Qualcomm port with 8 GiB DRAM using a 4 GiB owned window, measured on hardware. The failure is benign: the NO WINDOW path only prints the memory map on screen and writes no partition.
  • MemoryMapLib reconciliation (static, 2026-11-07, docs/step5-firmware.md section 'MemoryMapLib reconciliation (static)'; static analysis, not verified on hardware): the uefi-nuwa table (MemoryMapLib.c:9-35) has 41 entries, 27 inside the DDR range; the live FDT has 35 fixed-reg reserved regions, 6 of them no-map regions the table does not describe (xbl_ramdump_region 0x81200000+0x280000, global_sync_region 0x82600000+0x100000, mpss_dsm_region 0xd4d00000+0x3300000, xbl_sc_region 0xd8100000+0x40000, hwfence-shmem 0xe6440000+0x2dd000, llcc_lpi_region 0xff800000+0x600000). If RamManagerDxe pads every uncovered System RAM gap as EfiConventionalMemory, 16 fragments total about 10.80 GiB (largest 0x8c0000000 + about 9216 MiB ending at 0xafffff000, then 0x880000000 + about 925 MiB, 0xe7cb0000 + about 379 MiB, 0xbf300000 + about 346 MiB, 0x82800000 + about 128 MiB). A 1 GiB window therefore cannot fit any low fragment and a first-fit lands in the 0x8c0000000 fragment inside the ADT dram-base; shrinking owned_memory_size below about 925 MiB would move it to a low address outside the ADT range, so it must not be shrunk, and any memory-number change needs user confirmation. 'No reserved/gap overlap' is not safety evidence because System RAM already excludes no-map regions; the RAM Partition content is unknown.
  • Reviewer verdict on the first fastboot boot (2026-11-07, docs/step5-firmware.md section 'First fastboot boot: Reviewer verdict and conditions'; read-only static): the window-allocation risk is Medium rather than High, APPROVE WITH CONDITIONS scoped to the first fastboot boot only, and Deep Reviewer was not used because no candidate met the upgrade threshold in `../../docs/routing.md`. The failure path writes no partition (EFI_OUT_OF_RESOURCES at DarwinBoot.c:531-545; the Fail path prints boot refused and the on-screen NO WINDOW/FREE lines at :534-541), and the success path prints pa/va/entry/args/owned (:785-790, :797-811). XNU's physical range comes from boot_args (DarwinBoot.c:745-748), so the earlier 'window outside the ADT DRAM' premise is withdrawn; whether dram-base/dram-size has an XNU-side consumer is unconfirmed. The gEfiRamPartitionProtocolGuid producer stays unresolved (RamManagerDxe is not in apriori, Depex = TRUE, runs in FV order at nuwa.fdf:87).
  • Unresolved Medium hypothesis: MemoryMapLib was inherited from sheng and not reconciled item by item with nuwa; if the RAM Partition reports a no-map region that the table lacks as RAM, it would be padded as Conventional and used by DXE, giving a wrong access or reboot rather than flash corruption. The RAM Partition contents are unknown, so no conclusion is drawn.
  • Collect the first memory and dma observation, its cause when established, and the resulting implementation decision.

Completion criteria

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

Current limits

  • XNU handoff is withheld: the tail 4 KiB and the internal/dynamic reservations of the 9 GiB range are unclosed, and the owner of the 0xbe000000-0xd4c00000 reserved range that the DXE heap intersects is unknown.
  • No board DTS register set for DMA, storage, display, input and power yet, so those components stay unselected.
  • Window allocation is Medium risk, not High (static analysis, not verified on hardware): the gEfiRamPartitionProtocolGuid producer is unconfirmed and XNU's use of the ADT dram-base/dram-size is unconfirmed; the first boot will show NO WINDOW/FREE lines and write no partition if it fails. 6 live no-map regions are not in MemoryMapLib, and whether the RAM Partition reports them as RAM is unknown; adding them is a user decision. Do not shrink owned_memory_size: below about 925 MiB a first-fit window falls outside the ADT dram-base, and any memory-number change needs user confirmation.

Next action

  • Resolve the RAM Partition protocol producer (gEfiRamPartitionProtocolGuid) by binary reference-point analysis of InstallProtocolInterface versus LocateProtocol and by the first boot's DEBUG log, because RamManagerDxe's LocateProtocol failure is silent.
  • Decide with the user whether to add the 6 live no-map regions not described by MemoryMapLib to the static table; do not shrink owned_memory_size (a low first-fit window would fall outside the ADT dram-base).
  • First boot must use a DEBUG build and fastboot boot only (no partition write; device action needs user authorisation in that round), then read the DarwinBoot screen text and the DXE-stage AdvancedLogger page, and confirm SEC writes the entry parameters into the HOB (UniXNURamDiskDxe.c:215-233).
  • Close the High-memory gate: identify the active owner of the historical 0xbe000000-0xd4c00000 reserved range (DXE heap 0xbae00000-0xbf300000), account for the 0xafffff000-0xb00000000 tail and the internal reservations, then re-check the FD/ramoops overlap before allowing the XNU handoff.

Evidence

Storage and root

planned

Hardware and dependencies

  • Identify the storage and root hardware, resources and existing component interfaces for this board.

Configuration

  • Record the board.toml fields, DTS properties and defconfig values used for storage and root.

Procedure

  • Follow the storage and root stage in the porting guide and record the actual commands.

Development findings

  • 2026-11-15 milestone: the native UFS root volume mounts and initproc reaches user space. userdata's first 8446279680 B were written in 126 windows and verified by a whole-image readback (sha256 da429da64b4d91a08b8adb1d3162f8348b37b358509ab1bd5d8bf09abdbf68d7, installer exit 0, backup_policy=without-backup); the payload boots with native-ufs-root=1, native-ufs-lun=0, rd=uuid and boot-uuid DADD93F8-DBBF-4B17-A1DF-0F11E64172F7, Ramdisk_bytes 0. Device-side result is parent/user-reported (cyan band 3 = initproc exec finished, orange band 15 = initproc back in unix_syscall); the host artifacts are self-checked only. Authentication NOT REACHED, launchd completion/display/reboot persistence unverified. See docs/nuwa-launchd-reached-20261115.md.
  • 2026-11-15 bypass ledger for the root path: rootauth panic -> opt-in 8796-nuwa-root-auth-skip records nuwa_root_auth_error and continues; zero-initialised trust_cache_rt_lock -> opt-in 8796-nuwa-trust-cache-init restores trust_cache_runtime_init() + load_static_trust_cache() before machine_lockdown() (the ADT TrustCache memory-map entry is still &#123;0,0&#125;, so the load is a no-op); each carries an exit condition in docs/nuwa-launchd-reached-20261115.md. All three launchd-path patches stay opt-in and outside patches/xnu/series.
  • 2026-11-14 inspect-02 CRLF cause confirmed by original-capture-01: identical146B SHA c9450e5d55799e6d1ec14cfc24117f07ae343112d12a488188b617f81b073142, rc0/stderr empty, 13CRLF/0bareLF and complete CRLF footer. Minimal shell-text-only CRLF-to-LF parsing fix self-checked; full footer/ASCII/fields/rc gates preserved, call diagnostics retain raw bytes and exec-out GPT/data/image streams unchanged. Default/explicit waiver and all write gates retained. No device command or install this batch, no old whole matrix rerun. Older generic backup cause remains unknown. Evidence: out/nuwa-userdata-crlf-host-20261114/REPORT.md.
  • 2026-11-14 root observability r5 host-only self-checked: run-r5-visible-candidate kernel 55545856 B SHA256 809587d53fef9b0b3e9b2054a9e06c649d3c66cd45be3fafbe632ac575cb0acc. Root-only setup/publisher failure now terminates through existing panic renderer; runtime read failures use per-request transport-lock software snapshots and verify/copy evidence, panic after both locks release, and do not complete. Success completes exactly once. This supersedes r4 nine-failure-callback semantics, not its unchanged transport/quarantine policy. Five bounded root progress labels retain old marker IDs, 16-row capacity and framebuffer/cache publication; MEDIA_READY means signed UUID plus ancestor, not mounted. Actual native read/startRoot/publisher tests, final ARM64 marker execution and incremental helper-to-panic-to-renderer audit self-checked; independent verifier/review pending. No device, media write, packaging or mount/persistence evidence; fresh kernel-bound CHMH required. Evidence: out/nuwa-ufs-apfs-root-integration-20261114/observable-r5/REPORT.md.
  • 2026-11-14 latest user decision supersedes backup-first: explicitly install without another backup, not root boot. Host-only increment self-checked: default real-backup gate retained; --without-backup must join all existing write authorization/frozen-input gates and durably audits no backup/no rollback. Mandatory tools_inspect before write mkdir/push, same-boot/by-name/sysfs/GPT/offline/image/window/final-readback/claim gates unchanged. Only Windows adb.exe local push operand converted by actual host wslpath. Redacted stage/argv/rc/bounded stderr/original exception and stdout length/hash journaled before wrapping, including read-only failures and host open/fsync. Original failed backup cause unrecoverable, no guard/GPT attribution; cancelled output untouched and not reread. All device commands mocked, new review/verifier pending. Evidence: out/nuwa-userdata-no-backup-host-20261114/REPORT.md.
  • 2026-11-14 retained-domain UFS/APFS root integration host candidate self-checked: NativeUfs root=1/probe=0 initializes real storage classes, explicitly requires configured LUN0 and measured TUR/RC10 4K geometry, adopts retained 144KiB transport without production SMMU rewrite/fallback, binds GPT to measured capacity, preserves dual owner-policy and per-block verification, and publishes through existing IOStorage/APFS path. Continuous slot0 reads serialize through transport/window locks, completed tags may wrap, any transport/SCSI read failure permanently quarantines with IO completion. All media writes/store/log windows refused. Frozen run-r4 kernel 55545856 B SHA256 3ac37037104c98f11fc009abc42cf3495802f874ab3aa9fff1a929079454ad45; 920 native continuous/concurrent reads, 9 failure completions, incremental entry/geometry/GPT/policy-publication tests and final machine/APFS audit self-checked. No packaging/device/userdata change or mount/persistence proof; independent verifier/scout/deep pending. Evidence: out/nuwa-ufs-apfs-root-integration-20261114/REPORT.md.
  • 2026-11-14 root r4 CHMH/FAT/outer host packaging self-checked only; historical intermediate pending replacement by observable-failure kernel and fresh CHMH. Exact r4 output/negative tests and limits: out/nuwa-ufs-apfs-root-chmh-20261114-r1/REPORT.md. No root executor profile frozen, device action, userdata installation, or mount claim.
  • 2026-11-14 root r5 visible kernel rebound to CHMH/FAT/outer and isolated executor profile ufs-apfs-root-20261114-r2; actual outer readback, host-inputs process trap, three unauthorized dry phases and targeted negatives self-checked. Evidence out/nuwa-ufs-apfs-root-chmh-20261114-r2/REPORT.md. Independent packaging/executor review and verifier pending; no device action, userdata installation or mount claim.
  • 2026-11-14 combined LUN0 diagnostic isolated candidate: NOP, power Query, conditional single ACTIVE SSU to device WLUN, one TUR, RC10, one LBA0 READ10 with capacity-derived 512..131072 byte block; optional one REQUEST SENSE then terminate. Retained 144KiB full-span-qualified DMA, monotonic tags, completed-response ownership gate before reuse, no IS W1C, retry/reset/free/SMMU writes/media writes/root. Fixed accepted producer contract retained. Host-only self-checked; final independent verifier and directed deep pending, no device authorization. Evidence: out/nuwa-ufs-combined-read-20261114/RESULT.md.
  • 2026-11-14 preserved-NOP fixed-kernel CHMH packaging and isolated executor profile self-checked only: actual outer and four FAT members audited, fixed probe2/root0 args, terminal enum distinguished, no marker inferred. Experimental CHMH identity unauthenticated; kernel semantic deep review, packaging review and independent verifier pending. If kernel changes this package is withdrawn; no device operation or authorization. Evidence: out/nuwa-ufs-preserved-nop-chmh-20261114-r1/RESULT.md.
  • 2026-11-14 preserved-domain single-NOP real inspect candidate, self-checked only: immutable global once-claim, current SID/domain and full-span qualification, both queues bounded-stop/recheck, own transfer-base readback, CTR-derived publication/invalidate-only completion, one DB maximum and retained references at every endpoint. Fixed accepted r4 CPU producer contract consumed; no SMMU writes, task DB, reset/retry, root, packaging or device action. Host execution-template fault tests and final machine inventory complete; independent verifier and final directed deep pending. Evidence: out/nuwa-ufs-preserved-nop-exec-20261114/RESULT.md.
  • 2026-11-14 preserved-route host planner: measured SID60 slot2/bank2 mode words used only in synthetic complete snapshot; shared bank separately reported, widths explicitly supplied, submission always closed. No device ownership, identity-DMA or runtime qualification; self-checked only, independent review/verifier pending. Evidence: out/nuwa-ufs-preserved-route-plan-20261114/RESULT.md.
  • 2026-11-14 SID domain kernel host CHMH wrapper and isolated executor profile, self-checked only: exact probe2/root0 args, unauthenticated experimental identity, refusal and status4/status9 observations distinguished. No device authorization, ownership, translation or readiness claim; final directed deep and independent verifier pending. Evidence: out/nuwa-ufs-sid-domain-chmh-20261114-r1/RESULT.md.
  • 2026-11-14 SID0x60 bounded read-only domain candidate implemented in an isolated inspect-only overlay, self-checked: legal SMMUv2 4K geometry and EXIDS profile, complete bounded SMR uniqueness scan, unique translate bank CBAR/CBA2R/SCTLR mode/enable summary, explicit unsupported/nonunique/mapping refusal. All 256 mapping subpages PAR-qualified before MMIO; existing 16 marker rows and fMap lifetime reused. No TTBR/TCR decode, ownership claim, control write, ATOS, table/queue dereference, NOP, DMA, package, executor or device operation. Contract and pending directed deep/independent verifier: out/nuwa-ufs-sid-domain-20261114/CONTRACT.md.
  • 2026-11-14 queue-only five-read kernel increment independently verified; host CHMH packaging ordinary review and independent verifier PASS. Executor five-read profile self-checked only, independent review/verifier pending. HCE gates four queue-register reads, not an ownership or DMA claim. No device action. See docs/nuwa-ufs-queue-five-20261114.md, docs/nuwa-ufs-queue-five-chmh-20261114.md and docs/queue-five-executor-host-20261114.md.
  • 2026-11-14 handoff observation specification and synthetic correlation fixture self-checked only: same attempt/sequence, timer and EBS status, full allocation to map and kernel reservation coverage. No real observations or ownership proof; selected core producer-disabling path remains unestablished. TRACE corrected d2ec separate calls and post-scan unconditional rearm. No firmware, device, packaging or networking.
  • 2026-11-14 single-NOP r2 CLOSED, self-checked incremental only; original revision ordinary review PASS WITH CONDITIONS and independent verifier PASS at out/nuwa-ufs-nop-verifier-20261114/verifier-summary.md, not rerun. Added actual configure/start helper readback-failure partial-write trajectories, buffer binding/bytes retention and zero new doorbells on failure/reentry, plus successful NOP reentry refusal. Cache release is publication, not freeing allocation. Fixed OEM original/patched/INF hashes and selection reject altered input. Static trace out/nuwa-ufs-ownership-trace-20261114/TRACE.md closes OEM EBS registration-code callback target0x369c through CDB35/1b producer paths, timer cancel/rearm boundaries, shared list allocation origin, IOMMU protocol authority and pinned loader last SFS Close; registration success/execution, final BlockIO, old allocation lifetime and active SID domain still NOT OBSERVED. No new public source/device/packaging/root/GPT/media command; real entry/r7/executor unchanged. Directed deep only for concrete handoff blocker/implementation, not this host-only increment.
  • 2026-11-14 original single-NOP host work CLOSED; now independently reproduced for host scope only. Actual local atomic single-attempt exclusion, no hardware owner token; real entry refuses before MMIO/allocation. Pinned ShengUfs queue/SMMU/command/poll bodies reused in explicitly assumed host model; NativeUfs allocation/cache source preserved, not runtime-qualified. Nuwa live DT establishes SID0x60/apps-smmu, not active OEM domain. Selected OEM EBS registration code points to callback0x369c; successful registration/execution, producer revocation, old queue lifetimes and SMMU ownership unqualified. No packaging/device/root/GPT/media commands, official edit or r7/executor change.
  • 2026-11-14 HC executor profile ufs-hc-observe-20261114-r1 host-only self-checked, independent executor review and verifier pending; see docs/ufs-hc-executor-host-20261114.md. Parent reports r7 kernel deep and verifier PASS and packaging review/verifier PASS. No device authorization or action.
  • 2026-11-13 task A, self-checked: eleven IOBlockStorageDevice imports mapped independently in donor 2ec297ba and current r5 prelink by segment offset; class IOService parent/0x90 size/C2/D2 and 190 slots checked, four actual-byte mutations rejected and eleven offline aliases nm-checked. NativeUfs.o f673df62 preserved, actual iOS17 deployment retained for annotation reference built with the same 8796 flags. No final link, runtime/private-ABI acceptance, hardware or VerifiedRoot changes; layouts must be rebound. See docs/nuwa-ufs-host-20261113.md.
  • Launch route B decided 2026-11-06 (docs/decisions.md): bring-up runs from fastboot boot (route A, no partition writes), then the end state is the native UFS root with the userdata partition flashed as the root volume. B needs [userdata] LBA geometry from this board's GPT, root_match_volume_uuid from the built userdata image, native-ufs-root=1, CONFIG_XNU_NATIVE_UFS=y and the ShengUfs component; the flash overwrites Android data and requires explicit user authorisation.
  • Repo hygiene: removed the committed __pycache__ copies of the format-1 legacy tooling; no source behaviour changes.
  • Payload source decided 2026-11-07 (docs/decisions.md, subsection in section 1): route B keeps taking the boot payload (kernel, ADT, trustcache and unixnu.toml inside the RamDisk) from fastboot boot and does not flash boot.img into the boot partition. Observed: fastboot boot writes no partition, so the payload lives only for that boot. Inferred: every boot needs a host-supplied boot.img, and a device reboot returns to the Android boot partition instead of entering UniXNU, so the launch-route section 1 statement that a reboot goes straight into UniXNU and the host-free standalone boot end state do not hold today; standalone boot would need a separate decision plus explicit authorisation to flash boot. userdata remains the persistent root volume and flashing it still needs per-round authorisation; fastboot boot still needs a user confirmation before any device action.
  • Read-only GPT capture, 2026-11-07 (docs/device-evidence-2026-11-07.md): six LUNs sda..sdf are present and report 4096-byte logical and physical block sizes; userdata is sda32 with first_lba 3110976 and last_lba 61850618, 58739643 LBAs = 240597577728 bytes (about 224 GiB) at 4096 bytes per LBA, and its last_lba equals that disk's last usable LBA; peer entries are super sda31 751680..3110975, metadata sda17 16384..32767, persist sda18 32768..49151, boot_a sde14 149126..198277 (49152 LBAs = 201326592 bytes = 192 MiB), vendor_boot_a sde24 217670..242245, abl_a sde11, modem_a sde7 and xbl_a sdb1. board.toml [userdata] records the geometry; no GUID, serial or kernel command line is recorded.
  • Unit trap: GPT LBAs are 4096-byte units taken from the GPT header parsed on the host, while /sys/block/*/size and blockdev --getsz report 512-byte sectors, so the two must not be mixed without conversion (a factor of 8; for sda 61850624 x 4096 = 494804992 x 512). sda declares 32 GPT entries with usable LBAs 6..61850618. The recovery image does not support sgdisk -p (invalid option -p), so the GPT header and entry array were copied with dd to the host and parsed there; the device was only read.
  • Collect the first storage and root observation, its cause when established, and the resulting implementation decision.

Completion criteria

  • Host: register, clock, link and SMMU facts match the board inventory.
  • Reads: SCSI capacity, GPT and known blocks are correct; OS/SystemOS windows stay within userdata.
  • Root: the configured volume mounts and launchd starts. Persistence: writes survive reboot within the allocated Data region.

Current limits

  • The milestone is not the interaction target: authentication is NOT REACHED (unauthenticated experimental CHMH), launchd's own initialisation, the display service and any reboot-to-same-system behaviour are unverified, and the kernel reached user space only with three opt-in bypasses (root auth skipped, runtime trust cache empty, initproc exec policy ignored).
  • The userdata write and its whole-image readback are self-checked device-side: the installer reported exit 0 and the pinned hash, its independent verifier re-run (47 tests, hashes 1d8d077e/bb97229c/bc344eb1) was host-only with device_calls=0. The boot payloads in boot_a/vendor_boot_a have bootloader write acknowledgement only (no fastboot fetch, no adb in fastboot mode).
  • Hardware inputs and implementation remain incomplete: the board DTS has no measured storage/display register set, and the root volume's write persistence inside the allocated Data region is untested.
  • Any new kernel requires a freshly bound CHMH and cannot reuse the r7/r8/r9 packages; the observation kernel and the observe patches are temporary and must not be carried forward.

Next action

  • Next interaction work, after the 2026-11-15 milestone: verify that launchd finishes initialising (its exit/relaunch behaviour), land the display service through GopFramebuffer, then power-cycle once and check the device returns to the same persistent root. Keep the three bypasses opt-in and delete each one only when its exit condition is met (real trust cache, authenticated boot).
  • Boot payloads now live on the device: boot_a carries the Mu-Silicium loader plus DarwinBoot and vendor_boot_a the r9 package, both written under this round's authorisation, with rollback steps in out/nuwa-boot-payload-write-20261115-01/ROLLBACK-r9.md; keep that backup until the interaction target is reached.
  • Inventory the required storage and root inputs before implementation.
  • Flash timing confirmed 2026-11-07 (docs/decisions.md section 8): flashing userdata only validates persistence and comes after all first five bring-up stages, none of which writes a partition. Order: (1) firmware builds, memory ownership gate has evidence, and the boot-chain handoff passes independent review (after the Astra High channel recovers); (2) first fastboot boot observing UEFI, menu, payload and DarwinBoot, with the firmware log exported through the Tools QR path; (3) kernel entry through to platform expert; (4) read-only storage validation of the UFS controller, capacity, GPT and known data blocks, with the OS/SystemOS windows inside userdata; (5) build the root volume image, take the root UUID, check hashes and reach launchd on the in-memory boot path; (6) flash userdata and validate persistence. Before flashing, confirm item by item: the first five stages have real log evidence, the [userdata] geometry comes from a real read rather than an estimate, and the user has a usable stock fastboot package and confirms the Android data is no longer needed; the exact partitions, commands and impact are stated beforehand and executed only after explicit approval in that round. The boot partition is not flashed because the user decided to keep fastboot boot.

Evidence

Display

planned

Hardware and dependencies

  • Identify the display hardware, resources and existing component interfaces for this board.

Configuration

  • Record the board.toml fields, DTS properties and defconfig values used for display.

Procedure

  • Follow the display stage in the porting guide and record the actual commands.

Development findings

  • Collect the first display observation, its cause when established, and the resulting implementation decision.

Completion criteria

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

Current limits

  • Hardware inputs and implementation are pending for nuwa.

Next action

  • Inventory the required display inputs before implementation.

Evidence

Collect evidence using the next action.

Input

planned

Hardware and dependencies

  • Identify the input hardware, resources and existing component interfaces for this board.

Configuration

  • Record the board.toml fields, DTS properties and defconfig values used for input.

Procedure

  • Follow the input stage in the porting guide and record the actual commands.

Development findings

  • Collect the first input observation, its cause when established, and the resulting implementation decision.

Completion criteria

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

Current limits

  • Hardware inputs and implementation are pending for nuwa.

Next action

  • Inventory the required input inputs before implementation.

Evidence

Collect evidence using the next action.

GPU

planned

Hardware and dependencies

  • Identify the gpu hardware, resources and existing component interfaces for this board.

Configuration

  • Record the board.toml fields, DTS properties and defconfig values used for gpu.

Procedure

  • Follow the gpu stage in the porting guide and record the actual commands.

Development findings

  • Collect the first gpu observation, its cause when established, and the resulting implementation decision.

Completion criteria

  • GPU submissions complete, mapping/fence handling is correct, a known image renders, and the selected Metal client presents it.

Current limits

  • Hardware inputs and implementation are pending for nuwa.

Next action

  • Inventory the required gpu inputs before implementation.

Evidence

Collect evidence using the next action.

Network

planned

Hardware and dependencies

  • Identify the network hardware, resources and existing component interfaces for this board.

Configuration

  • Record the board.toml fields, DTS properties and defconfig values used for network.

Procedure

  • Follow the network stage in the porting guide and record the actual commands.

Development findings

  • Collect the first network observation, its cause when established, and the resulting implementation decision.

Completion criteria

  • Scanning, association, address configuration and sustained receive/transmit traffic pass with the selected firmware.

Current limits

  • Hardware inputs and implementation are pending for nuwa.

Next action

  • Inventory the required network inputs before implementation.

Evidence

Collect evidence using the next action.

Audio

planned

Hardware and dependencies

  • Identify the audio hardware, resources and existing component interfaces for this board.

Configuration

  • Record the board.toml fields, DTS properties and defconfig values used for audio.

Procedure

  • Follow the audio stage in the porting guide and record the actual commands.

Development findings

  • Collect the first audio observation, its cause when established, and the resulting implementation decision.

Completion criteria

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

Current limits

  • Hardware inputs and implementation are pending for nuwa.

Next action

  • Inventory the required audio inputs before implementation.

Evidence

Collect evidence using the next action.

Power

planned

Hardware and dependencies

  • Identify the power hardware, resources and existing component interfaces for this board.

Configuration

  • Record the board.toml fields, DTS properties and defconfig values used for power.

Procedure

  • /home/wh1stl3/.local/bin/python3.11 -B scripts/build-nuwa-pmgr-kernel.py --out out/nuwa-pmgr-kernel-20261113-r5 --reuse-provider out/nuwa-pmgr-kernel-20261113-r2/provider
  • /home/wh1stl3/.local/bin/python3.11 -B scripts/audit-nuwa-pmgr-kernel.py --build out/nuwa-pmgr-kernel-20261113-r5 --python-tools /home/wh1stl3/.local/lib/python3.10/site-packages

Development findings

  • 2026-11-15 the device source tree now carries package/NuwaPlatformPower.cpp and package/diagnostics/* unchanged from the reviewed r5/r2 sources; the r5 kernel that reached launchd boots with nuwa-pmgr=2, but no power hardware, real nub publication or AMFI consumption is claimed by landing the source.
  • 2026-11-13 executor profile pmgr-mode2-20261113-r1 self-checked only; Windows and WSL real host gates, 115 synthetic groups, 18 unauthorised CLI dry calls, zero actual device commands; independent executor review and verifier pending. Parent reports kernel directed deep review and verifier PASS, and package ordinary review and verifier PASS.
  • Historical r2 guarded-provider archive (568885fd698489843701907e5d8216f3c9b377c678a2d6710bd8652bad93d29e) retains its ordinary review and independent verifier evidence; source now takes a borrowed root-publication parent, so that acceptance is not claimed for the new link. Original ApplePMGR storage 0x36968, 325 vslots, nub 0x740 and nub+0x88 provider binding remain; unsupported power slots refuse, real readAon/writeAon retained, no hardware/MMIO success claim.
  • 2026-11-13 isolated kernel delivered, self-checked: out/nuwa-pmgr-kernel-20261113-r5/kernel.macho 55451648 B SHA256 81b85228316574782f9d5390956c3ce3d8c97e4b345a9ea1631feacf3280b5d1. Parent supplied ordinary review PASS WITH CONDITIONS for explicit root-parent route, superseding the initial no-platform blocker. Device-owned opt-in patch calls nuwa_pmgr_run(this) once in IOPlatformExpertDevice::startIOServiceMatching after workloop success, before original registerService; no global synchronous matching or AMFI edit. Native relink, donor pthread prelink, original renderer sources/font and other native inputs reused with provenance. Eight donor/segment/slide bindings rederived and final-checked, all 325 slots final-checked including native inheritances, initializer manual only/no auto-init. Startup audit asserts Sandbox GOT -> AMFI policyops+0x50 -> later BSD callback -> kernel_thread_start(worker): hook precedes worker creation statically. Ten bounded real-code checks pass; mode1 cleanup/mode2 publication/full mode3 lifecycle and pthread runtime unexecuted. New kernel requires fresh CHMH and directed deep review/independent verifier; no device release.

Completion criteria

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

Current limits

  • Hardware inputs and implementation are pending for nuwa.
  • The guarded provider is a host build only: publication and power-refusal lifecycle are not runtime-qualified, and no power hardware is claimed. The system llvm-objcopy (LLVM 14) does not support Mach-O --remove-section/--globalize-symbol, so the build script performs the symbol move and __mod_init_func disable in place and re-verifies the partition invariants and __text equality. Independent r2 rebuilds changed both __debug_line and __debug_info and section/symbol offsets: debug paths can affect debug content/layout, with no fixed-size guarantee. Compare non-debug contents plus relocation semantics and the eight bindings, not full archive hashes.

Next action

  • Parent reports kernel directed deep review PASS WITH CONDITIONS and independent verifier PASS, and mode2 CHMH package ordinary review and independent verifier PASS. New executor profile is self-checked only; independent review and verifier pending. The target is actual nub publication and AMFI consumption, not merely disappearance of the old panic. Software PMGR is not hardware; CHMH identity is unauthenticated and authentication NOT REACHED. No device action authorized. Details: docs/pmgr-executor-host-20261113.md.

Evidence

Image assembly

planned

Hardware and dependencies

  • Identify the image assembly hardware, resources and existing component interfaces for this board.

Configuration

  • 2026-11-10 current candidate on 982e92e67c0e266987a786865f53e9cd120a159c plus this board/docs group: kernel_args=-v serial=3 sheng-platform-stage=5, panic_screen=true, entry_probe_reset_delay=0, old reset selectors removed. Existing ArmDTPlatform.cpp:191-208 resets immediately after registerService returns, not after six seconds; no sheng-platform-panic argument. Source prelink remains 55357440 B SHA256 da0d660f8336fa87825ec34ce1f48b653d728a6277aaf7d4b03784cf4f73c969; kernel/prelink not rebuilt. PAN and volatile fallback evidence is retained in docs/panic-fallback-host-20261110.md. Crypto correction: CONFIG_XNU_CRYPTO_BOOTSTRAP=prelinked, prng_random.o final -U XNU_SOURCE_CORECRYPTO14, no _source_corecrypto14_bootstrap in current prelink; console return does not imply a subsequent source-corecrypto EARLY_BOOT path. Mode unchanged.
  • Single isolated ext4 assembly rc=0; parent-delivered independent Verifier incremental audit PASS, rc=0. Renderer 0xa0f bytes at 0xfffffff0073ee5c0, end 0xfffffff0073eefcf, unchanged boundary 0xfffffff0073eeff0, spare 0x21. Console site 0xfffffff0079993bc restored BL 0xfffffff007968f5c, no stage3/4 reset. Ordinary SP and common va_list* entries checked; manifests/hash/ramdisk/partition size match. Full evidence and local output paths: docs/stage5-panic-host-20261110.md. Previous measured candidate preserved (not rehashed by Verifier this pass). New candidate not device-tested. Historical findings/evidence below are dated snapshots, not current blockers or configuration.

Procedure

  • WSL Ubuntu-22.04, cwd /mnt/d/Code/PortingiOS/UniXNU: bash ../nuwa/UniXNU-nuwa/.cache/run-stage5-panic-host-20261110.sh; one payload/vendor_boot assembly, both rc=0, ext4 --out and existing --kernel. Exact expanded commands are in docs/stage5-panic-host-20261110.md and local assembly.log.
  • bash ../nuwa/UniXNU-nuwa/.cache/run-stage5-panic-check-20261110.sh: final audit rc=0, self-checked; no kernel/prelink rebuild or device action. Reuse unchanged ADT/trust evidence. Verifier should rerun the read-only audit with a new output directory, record check and git diff --check, not the assembly driver.

Development findings

  • 2026-11-15 launchd-path packaging: 8796-nuwa-root-auth-skip, 8796-nuwa-trust-cache-init and 8796-nuwa-exec-policy-skip are landed as opt-in device patches (applied with -DXNU_NUWA_ROOT_AUTH_SKIP=1 / -DXNU_NUWA_TRUST_CACHE_INIT=1 / -DXNU_NUWA_EXEC_POLICY_SKIP=1) and deliberately stay outside patches/xnu/series, the same convention as 8796-nuwa-pmgr-root-publication; r7/r8/r9 vendor_boot packages (100663296 B each) were built from kernels 2561c0c5/b1029f3d and written to vendor_boot_a. The observation patches (8796-nuwa-observe-markers, 8796-nuwa-observe-log) remain untracked pending the observation cleanup.
  • 2026-11-15 trust cache bundle: scripts/build-ios16-system-trust-bundle.py re-emits the two unchanged 20H392 modules (StaticTrustCache 2760 entries and Cryptex1,SystemTrustCache 61 entries) as the loader-accepted bundle eb1475f643f45a974ebd40596f6e2b1cdf40205d2ca3c6094d774886099399c3 (67764 B, no cdhash rewritten), with ten parser negatives recorded. board.toml [payload] trust_cache_sha256 still pins the older single-module 367e051548f9f0407fa9fe9d40007fa4e7c5ff27c45f5a7a30cedf344e24d98e that r3-r7 and the r8obs control package carried; the r8/r9 payloads used the two-module bundle.
  • Repo hygiene: removed the committed __pycache__ copies of the format-1 legacy tooling; no source behaviour changes.
  • Format-2 migration: scripts/package.sh taken unmodified from the official template; patches/xnu/series created empty; package/ documents the empty device package.
  • Kernel patch set (2026-11-08, host-only, self-checked): patches/xnu/series now lists the 17-entry minimal set for stage 5 and carries four local 8796 ports (8796-apple-only-cpu-tunables, 8796-non-apple-iommu-bootstrap, 8796-source-machine-layer, 8796-zzx-pmap-sync-64bit-pa); a static review found two build blockers in the first versions (a dropped ptep_get_iommu hunk, and a kpc source file that 8796 does not ship), fixed in r2. Update (2026-11-08, host build, self-checked): the series grew to 22 entries with six more local 8796 fixes found by real build rounds on the Linux toolchain: 8796-unifdef-headers (two headers the host unifdef 2.10 rejects), 8796-clang-var-init-flag (drops a legacy clang flag and silences new clang-24 warnings), 8796-indirection (os_get_pointee_type for the kalloc macros), 8796-kperf-kpc-off (guards kperf/kpc references because 8796 has no osfmk/arm64/kpc.c), 8796-uio-stackbuf (UIO_STACKBUF macro, 36 call sites). Update 2 (2026-11-08, host build, self-checked): the series is now 25 entries; the kernel compiles and links on the Linux toolchain, the archive steps (corecrypto-14 twice, ArmDTPlatform, ArmCpu) pass, and the final kernel is mach.development.bcm2837.unstripped (17281280 bytes, Mach-O arm64, bootstrap-pass link). Added since: 8796-iorpc-kernel-content (the host iig emits 10063-style code that uses IORPC.kernelContent, an additive member under KERNEL), 8796-kperf-kpc-link-guards (six undefined kperf/kpc symbols) and 8796-compile-records (10063's MakeInc.rule writes <obj>.o.json compile records with -MJ, 8796 does not, and the archive builders read them). The kernel archive list in board.toml is replaced with the four entries stage 5 uses (the arch list also names SoftwareAES, its AES bindings, mockfs and the source libpthread-519, whose pthread_callbacks_s layout does not match 8796), with CONFIG_XNU_NATIVE_AES=n and CONFIG_XNU_SOURCE_PTHREAD=n in nuwa_defconfig. The prelink step now runs and stops on the first ABI gap against the donor kernelcache: kernel target 0xfffffff00734c520 (_vfs_getvfs_with_vfsops) has no native address candidate; more gaps are likely. The IORPC patch and the other 8796 fixes are not reviewed beyond the first four ports. Update 3 (2026-11-08, host build, self-checked): the series is 26 entries; 8796-sparse-prelink-index.patch is the XnuPrelink kit patch xnu-10063.141.1-sparse-index.patch converted to LF (the kit patches are checked out with CRLF line endings, which is why git apply failed on them), touching only osfmk/kern/kalloc.c kalloc_type_view_parse (an early-boot allocator path; not reviewed). The full build applied all 26 patches and linked a 17315480-byte kernel. board.toml now drops com.apple.filesystems.apfs (it references _vfs_getvfs_with_vfsops, absent from open-source 8796; the 10063 definition is a short mount-list lookup) and com.apple.kec.Libm (a literal pool the planner reads as code) from [kernel.prelink] roots, 29 of 31 roots remain, and the planner then passes. Prelink now stops later, in emit_prelink.py instruction_inventory: External PC-relative encoding at 0xfffffff005b8c1cc (a kext in the remaining roots, not yet identified). The security-init kit patch (needed once early_security_init is set) also applies to 8796 after the same LF conversion. No prelinked kernel exists yet.
  • Collect the first image assembly observation, its cause when established, and the resulting implementation decision.
  • Payload built host-only (self-checked): ./unixnu build payload -D ../nuwa/UniXNU-nuwa --trust-cache out/ipsw/20H392/087-86698-044.dmg.trustcache (no --data-ramdisk) exited 0; payload.img 67,108,864 bytes (64 MiB), sha256 7bbffac545106c8710108737a61af3d15a0f86a5534cbc4d3bae9ee6aac958da. FAT16 volume NUWAESP carries EFI/UniXNU/unixnu.toml (402 B) and Payloads/&#123;kernel.macho 55,357,440 B sha256 274b5033..., kernel.macho.adt 13,132 B sha256 b7ac4b9c..., kernel.macho.trustcache 66,289 B&#125;; the ADT sidecar carries product-id/artwork-device-subtype/compatible-device-fallback, chip-id 4acab8db and the unique-chip-id seed (from the git-ignored unixnu.local.toml).
  • vendor_boot repacked host-only (self-checked): ./unixnu vendor_boot --stock UniXNU-nuwa/.cache/stock/vendor_boot_b.img --ramdisk payload.img -o out/nuwa/ios-16.7.16-20H392/nuwa/vendor_boot/vendor_boot.img exited 0; 100,663,296 bytes (96 MiB = partition), sha256 a6d2cc58beac8d89b9b9fd46e8de409374073899ae72440be6df32907717164f, vendor ramdisk 67,108,864 bytes at the stock address 0x01000000. The >15 MiB --ramdisk-addr relocation (docs/zh/guide/boot.md) is left at stock: the nuwa ABL relocates the initrd to the Kernel region (~0xB7xxxxxx per docs/step5-firmware.md) and no safe relocation address is established, so it stays a pending user decision before the first device boot.
  • Trust cache repack (2026-11-09, host-only, self-checked): the first payload's kernel.macho.trustcache was the raw StaticTrustCache IM4P (66,289 B, DER 0x30 0x82, sha256 63e01c03…) copied verbatim by 'build payload --trust-cache <im4p>'; ValidTrustCaches rejects it (its u32 Count reads 0x02018230 > 128), returning EFI_COMPROMISED_DATA when the xnu menu entry was selected on hardware. Re-bundling the same donor module with scripts/bundle-ipsw-trust.py yields a valid bundle (66,272 B, Count=1, offset 8, version-2 module, 2760 entries, sha256 367e0515…) that passes the loader-equivalent structural check; 'build payload --trust-cache <bundle>' then 'vendor_boot' repack succeed, and the rebuilt payload.img (sha256 b0c808c2…) / vendor_boot.img (sha256 fef65a4f…) carry the bundle.
  • Entry-probe payload build (2026-11-09, host-only, self-checked): ./unixnu build payload -D ../nuwa/UniXNU-nuwa --uefi ../nuwa/uefi-nuwa --trust-cache out/ipsw/20H392/087-86698-044.dmg.trustcache.bundle exited 0; kernel.macho copy patched 274b5033... -> 63b6ff91... (entry 0xfffffff0073ee488 b start_first_cpu -> b cave1; MMU-enable 0xfffffff0073f0350 msr sctlr_el1,x0 -> b cave2; _panic 0xfffffff00798d088 -> b panic_screen; _panic_trap_to_debugger 0xfffffff00726c174 -> b panic_screen_common); payload.img 67108864 B (64 MiB < 96 MiB partition) sha256 c59bc2bc02d58b3877e38d4af4d9f3ee1c44457ced1d09252d1fee9f8982fc98; ./unixnu vendor_boot --stock .cache/stock/vendor_boot_b.img --ramdisk payload.img exited 0, vendor_boot.img 100663296 B sha256 d3dc18f359a99373fa57d80b482138cf4834a51b5f71324f8ef50b999a61eb18; ADT sha256 varies per build (devicetree.generate writes os.urandom(256) random-seed).
  • Stage 4 hardware observation (2026-11-09, parent session, device): DarwinBoot validation passed and handed off to the kernel entry (entry words 0x140006f5 = b start_first_cpu); the screen then stays on the green handoff bar with no kernel output. Read-out: graphics mode (boot_args.Video[1] v_display = 1, hardcoded BootVideo.h:31) suppresses early printf, so no text does not prove a hang; -v is not parsed by iOS arm64 xnu; the boot_args CommandLine ABI size mismatch (608-byte Darwin 22 vs 1024-byte Darwin 23) is a known item. The entry-probe payload (blue/magenta bars) is the next observation step.
  • Entry probe payload (2026-11-09, host build, self-checked): ./unixnu build payload -D ../nuwa/UniXNU-nuwa --uefi ../nuwa/uefi-nuwa --trust-cache out/ipsw/20H392/087-86698-044.dmg.trustcache.bundle (exit 0) patches a copy of kernel.macho with scripts/patch-xnu-entry-probe.py --reset-delay 6 (no --panic-screen): stage 1 blue bar at the MMU-off entry, stage 2 magenta bar just before msr SCTLR_EL1, stage 3 busy-waits 0x06DDD000 = 115200000 ticks (19.2 MHz x 6 s) on CNTPCT_EL0 then x0 = 0x84000009 / smc #0. Disassembly verified the three branch-to-cave sites and that _panic/_panic_trap_to_debugger are not redirected; panic_screen stays false because panic-screen + stage 3 overflow the padding past PANIC_CAVE_END by 95 B (see the configuration area). payload.img 67108864 B sha256 dd54e153aa2c5a779d4d260493c0561c4c249fc80c3374781444f3714824060d (payload_kernel_sha256 e4b2f2d2f8fe357c18b616a4bb3253318d18fa4e95afc05d1fef3238406559ee vs kernel_sha256 274b50330e8d945480eff24e759fbe3b64d2329c002d78d29b405ad156aac3f0); ./unixnu vendor_boot --stock ../nuwa/UniXNU-nuwa/.cache/stock/vendor_boot_b.img --ramdisk <payload> -o <vendor_boot> (exit 0) vendor_boot.img 100663296 B sha256 7e6e519263323356de5e237cd7681c3233ae4ddad1b3544a597f93bc0f8502d5 (ramdisk 67108864 B at 0x1000, address 0x1000000). Both under the 96 MiB vendor_boot partition.
  • Entry-probe stage-4 bisect payload (2026-11-09, host build, self-checked): ./unixnu build payload -D ../nuwa/UniXNU-nuwa --uefi ../nuwa/uefi-nuwa --trust-cache out/ipsw/20H392/087-86698-044.dmg.trustcache.bundle (exit 0) patches a copy of kernel.macho with scripts/patch-xnu-entry-probe.py --reset-delay 6 --reset-after _ml_parse_cpu_topology --in _arm_init --occurrence 1; the bl at 0xfffffff007998914 is rewritten to b 0xfffffff0073ee620 (cave 4), whose code re-runs bl _ml_parse_cpu_topology, busy-waits 0x06DDD000 = 115200000 ticks (19.2 MHz x 6 s) on CNTPCT_EL0 then x0 = 0x84000009 / smc #0 followed by a wfe loop. payload.img 67108864 B (64 MiB < 96 MiB) sha256 e4cb58c429b90cf678222c37c8468820094225b805fcdfb541f12306aa9fd24b (payload_kernel_sha256 18d7668ec271babff0e3a7832523551bbac96d454101a35fa30c72955627b8c4 vs kernel_sha256 274b50330e8d945480eff24e759fbe3b64d2329c002d78d29b405ad156aac3f0); FAT16 NUWAESP Payloads/&#123;kernel.macho 55357440 B, kernel.macho.adt 13132 B sha256 5066738a..., kernel.macho.trustcache 66272 B sha256 367e0515... bundle Count=1 version-2&#125;; ./unixnu vendor_boot --stock ../nuwa/UniXNU-nuwa/.cache/stock/vendor_boot_b.img --ramdisk <payload> -o <vendor_boot> (exit 0) vendor_boot.img 100663296 B sha256 81580c186275bc0e87d17a97ddb3565407093344d13155a3b415d1859a99f298 (ramdisk 67108864 B at 0x1000, address 0x1000000); vendor_boot ramdisk region sha256 == payload.img sha256. ADT sha256 varies per build (devicetree.generate writes os.urandom(256) random-seed).
  • Stage-4 next probe (2026-11-09, host-only candidate, self-checked): user-reported hardware reset after about 6 s with revision 07e176761020d3d7f01fb086d9a89af3e3531eb9 proves _ml_parse_cpu_topology returned. board.toml now selects _arm_vm_init in _arm_init, occurrence 1, delay 6, panic_screen=false; reset means arm_vm_init returned, no reset is inconclusive (not reached, did not return, or probe/reset failure). Cave 4 at 0xfffffff0073ee620 is inside kernel __TEXT_EXEC; patch-xnu-entry-probe.py:97-118 uses PC-relative BL/literal load, a 115200000-tick constant, register-only CNTPCT_EL0/PSCI SMC. xnu-8796.141.3 osfmk/arm64/arm_vm_init.c:1998 obtains the KVA segment, :1413 maps the full __TEXT_EXEC ROX, :2097 calls arm_vm_prot_init, and :2144-2156 switches tables and invalidates the bootstrap identity map. Identity-map removal alone therefore does not invalidate the cave; actual post-return execution remains untested. Existing payload/vendor_boot image and JSON .v5 backups were retained without overwrite. Both build commands in the packaging procedure exited 0; the payload is 64 MiB <96 MiB, while stock-padded vendor_boot is exactly 96 MiB, NOT strictly smaller. No device action or push. Final audit exited 0: call 0xfffffff007998a3c -> cave 0xfffffff0073ee620 -> BL _arm_vm_init; CNTPCT loop, literal 115200000, PSCI 0x84000009 / smc #0; blue/magenta sites retained, 0xfffffff0073f0370 and both panic entries unchanged. FAT32 (not the older record's FAT16) contains kernel 55357440 B, ADT 13132 B, trustcache bundle 66272 B (Count=1, offset=8, version=2, entries=2760). Payload SHA256 5253c193a39daf2e663fd98496299894bbfa7d6f7ba0c98c14c5b625d2a922df; stock-padded vendor_boot SHA256 4f741c558ccd80211db6bca6cd8098d9efe825a3409902b775369a01788c65b1. Vendor ramdisk region SHA256 equals payload SHA256. Commands: ./unixnu build payload -D ../nuwa/UniXNU-nuwa --uefi ../nuwa/uefi-nuwa --trust-cache out/ipsw/20H392/087-86698-044.dmg.trustcache.bundle; ./unixnu vendor_boot --stock ../nuwa/UniXNU-nuwa/.cache/stock/vendor_boot_b.img --ramdisk out/nuwa/ios-16.7.16-20H392/nuwa/payload/payload.img -o out/nuwa/ios-16.7.16-20H392/nuwa/vendor_boot/vendor_boot.img. Independent verification pending.
  • Host commands (WSL Ubuntu-22.04, cwd /mnt/d/Code/PortingiOS/UniXNU, PATH=$HOME/.local/bin:/usr/local/bin:/usr/bin:/bin; device base revision 2d55e2a8d223f558f01c948db45db3c71f747c96 plus this board.toml change): ./unixnu build payload -D ../nuwa/UniXNU-nuwa --uefi ../nuwa/uefi-nuwa --trust-cache out/ipsw/20H392/087-86698-044.dmg.trustcache.bundle; ./unixnu vendor_boot --stock ../nuwa/UniXNU-nuwa/.cache/stock/vendor_boot_b.img --ramdisk out/nuwa/ios-16.7.16-20H392/nuwa/payload/payload.img -o out/nuwa/ios-16.7.16-20H392/nuwa/vendor_boot/vendor_boot.img. Both rc=0, self-checked. Old payload/vendor_boot image and JSON .v6 backups were absent and then copied and cmp-checked without overwrite. Final FAT extraction and llvm-objdump audit (python3 -u /mnt/c/Users/Administrator/AppData/Local/Temp/opencode/nuwa-console-audit.py, rc=0): 0xfffffff007998b84 B 0xfffffff0073ee620; cave BL _PE_create_console at 0xfffffff00796865c, CNTPCT_EL0 loop with literal 115200000 at 0xfffffff0073ee660, x0=0x84000009 then SMC #0 (word 0xd4000003 at 0xfffffff0073ee654). Original _arm_vm_init BL at 0xfffffff007998a3c and topology BL at 0xfffffff007998914 restored; blue/magenta branches retained; stage-3 site 0xfffffff0073f0370 and both panic entries unchanged. Extracted kernel 55357440 B SHA256 c62bac8e32000ada42eedefdd80ff56ddef1dcea3d441728b37b149abcb3e2d6; trustcache 66272 B SHA256 367e051548f9f0407fa9fe9d40007fa4e7c5ff27c45f5a7a30cedf344e24d98e. Local out/nuwa/ios-16.7.16-20H392/nuwa/payload/payload.img is 67108864 B SHA256 982a1d3b1234c3e461c5a1109fee3de0c95cab13eefa20841b606bbe6ce13620; vendor_boot/vendor_boot.img is 100663296 B SHA256 cc8f6712cb269ded26dd7b4f45aec1e26d90683e2cbf52b8b3fac02a6f651a7d. Vendor ramdisk is byte-for-byte equal to payload; 64 MiB payload and stock-padded 96 MiB vendor image are <=96 MiB partition. Payload hashes may change on rebuild because ADT random-seed is generated afresh. Independent verifier has not rerun these commands.

Completion criteria

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

Current limits

  • Hardware inputs and implementation are pending for nuwa.

Next action

  • Data-ledger integration still requires applicable static review; mechanical/schema completion is not authentication PASS or expanded5152 admission. Other authorized local implementation/validation, recovery-tool repair and independent firmware build continue within their own scopes. Before device use confirm authorization; RAM smoke->super-only windowed write/readback already has this-round authorization, not withdrawn here. Current0 byte stop is bootmode unknown, no background writer; general repairs compatibility, original25tests do not validate that pending fix. This document update ran no research/device command and changed no script/recovery tool; that is task-local history, not a global restriction. Retain host8ae344cc05879a27cf77b5103e30b2ca51a6d41d/firmware719afa67e3aebbf058c2a8beff4c801057fb2b20 and BNCH gate. userdata Android, iOS UFS absent, future iOS restoration requires fresh authorization. No further verification disabling/locking, public read-only/no external messages. Documents uncommitted; identity/target/readback gates unchanged.

Evidence

Diagnostics

planned

Hardware and dependencies

  • Identify the diagnostics hardware, resources and existing component interfaces for this board.

Configuration

  • Record the board.toml fields, DTS properties and defconfig values used for diagnostics.

Procedure

  • Follow the diagnostics stage in the porting guide and record the actual commands.

Development findings

  • 2026-11-15 observation markers (temporary, host only): the opt-in XNU_NUWA_OBSERVE build in out/nuwa-observe-20261115-r1 paints cyan band 3 from __mac_execve when process_signature returned 0 for initproc and orange band 15 at the first unix_syscall() entry with proc == initproc; bounded stores into vinfo.v_baseaddr only (no lock, allocation, sleep or call-out), idempotent per band, bands chosen so they never overlap the loader text block or the panic screen clear. Build/selfcheck are host-side; the two patches (8796-nuwa-observe-markers.patch, 8796-nuwa-observe-log.patch) stay untracked until the observation round is cleaned up.
  • unixnu.lock committed from the first complete fetch of the iOS 16 profile: xnu 1b191cb5 (8796.141.3), sm8550 a9f45dc6, edk2 cfb93724, libplatform a00a4cc3, libdispatch 30fe6a7d, libpthread c0e60fb5, AvailabilityVersions 933f9360, corecrypto-2023 11fba65b, IOStorageFamily e45fbe30.
  • The Tools entry is enabled in the payload configuration ([payload] tools = true in board.toml); the firmware log exporter now has a producer (Advanced Logger wired in uefi-nuwa e494c01b2f679439fdaeaf2084b693c7220543f5, built; independently re-run by the Verifier subagent on b1064eb8/e494c01b), but the Tools -> Advanced Logger page and its runtime dispatch are unverified and the entry has not been exercised on hardware.
  • Step 5 platform facts are recorded in docs/step5-firmware.md with provenance only: the backups MANIFEST is read before any other file, then recovery-cpu-20261021T005937467055Z/&#123;running-fdt.dts,iomem.txt,cpuinfo.txt&#125;, legacy-nuwa-artifacts/platform-inputs/20261020T124109389461Z/nuwa-platform-observed.json and memory-layout/nuwa-layout-tmlq93uo/report.json. No raw dump, device identifier, serial or seed is copied, and the archives were not re-scanned this round.
  • This review ran no host command and touched no device, so it adds no host or device evidence; the report is a source-level record of an independent review (UniXNU 693f5d9cae8a408dd18b39f3f9fdc4228c934241, device 279c990a0c7a238aa0a5818493c6cd6a46528749, SoC a9f45dc60c035084b95463781538380f4c3606c9, uefi-nuwa 322d742fd87d370aac96bb394d1feac61beb9b3a).
  • Recovery reader confirmed read-only on 2026-11-07 (docs/device-evidence-2026-11-07.md): a TWRP 3.7.1_12 recovery with a root adb shell provides adb, dd, cat of /proc and /sys, dmesg and blockdev --get*, so a recovery+adb log reader is available in practice; the only tooling gap seen was sgdisk -p, which that image does not support. The capture wrote no partition and ran no fastboot command.
  • pstore backend observed: /dev/pmsg0 exists while /sys/fs/pstore is empty, so the Android kernel has a ramoops backend; whether ramoops contents survive a warm reboot was not tested, and the ramoops address range is the one whose overlap with the FD window is recorded in the memory area as an observation only.
  • Collect the first diagnostics observation, its cause when established, and the resulting implementation decision.
  • 2026-11-09, user-reported hardware evidence relayed by the parent session: the six-second _arm_vm_init post-return probe at device revision 2d55e2a8d223f558f01c948db45db3c71f747c96 reset on hardware. This establishes that callee returned; it does not establish that subsequent code must hang. The implementer ran no device command this round. Next diagnostic candidate (2026-11-09, host-only, self-checked; Reviewer PASS WITH CONDITIONS supplied by parent): board.toml [payload] panic_screen=false, entry_probe_reset_delay=6, entry_probe_reset_after=_PE_create_console, entry_probe_reset_in=_arm_init, entry_probe_reset_occurrence=1, entry_probe_reset_before=false. The actual BL site is 0xfffffff007998b84; cave 0xfffffff0073ee620 lies in __TEXT_EXEC RX. _PE_create_console tail-calls initialize_screen, which can return. The six-second delay starts after the callee returns, not at boot or at the magenta bar. A probe reset proves only this callee returned, not that the following PE_init_printf ran; no reset is inconclusive (not reached, did not return, or probe/reset failure). The BCM2837 fixed panic-buffer access at 0x3c0fc000 in pexpert/arm/pe_init.c remains a risk and is not changed. No VM, memory-window, console implementation or base-source change, no device operation and no push; all area statuses remain planned. This finding supersedes historical claims that no reset necessarily means no return. New candidate not hardware-tested; independent verification pending.
  • 2026-11-11 device checkpoint (parent-delivered review; no device command rerun for this records update): the frozen pmap-ret-stop-20261111-r1 candidate was accepted on the host (outer static review PASS WITH CONDITIONS, independent verifier PASS, range static PASS) and, under explicit round authorization, scripts/run-pmap-ret-once.py ran one three-phase attempt (run pmap-ret-20261111-r4). Stage 2 command review PASS: the A-to-B switch read back B and the single fastboot boot of the fixed r4 loader exited 0. Stage 3 review PASS after an independent recomputation of the host readback: the device switched back to A and read back A, Android reported boot_completed=1/_a/uid0, vendor_boot_b was restored to this round stock backup 100663296 B SHA256 493ca426964adabb195a00ded4ff49793bc7efd5d7bea2a567bcfd0c970a8dad with the whole-partition hash equal to restore-readback.img, the r4 staging files were removed, and the stock backup and evidence were retained. The user reports the on-screen line DIAG: PMAP initializer normal epilogue reached; STOP; no recording file exists for independent visual confirmation, so this does not prove the original RET executed and is not authentication, stage5 or launchd evidence. Diagnostics status remains planned.

Completion criteria

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

Current limits

  • The Tools entry is configured and the Advanced Logger producer is wired (uefi-nuwa e494c01b2f679439fdaeaf2084b693c7220543f5); it was built and independently re-run by the Verifier subagent on b1064eb8/e494c01b; run-time unverified, so the Tools -> Advanced Logger page is still unverified at runtime, the 4 MiB buffer placement is an assumption and the rework has no independent boot-chain review.
  • The recorded platform facts are source-level only and carry no host or device verification.
  • Persistent kernel logging is unavailable until the route B ShengUfs storage driver works: the log region is a reservation inside userdata, and although the [userdata] GPT geometry is now measured, no region has been reserved and no writer or reader is implemented.
  • The log reader is still unselected between a minimal Linux image and recovery+adb, but the recovery option is no longer unknown: TWRP 3.7.1_12 with root, adb, dd and blockdev reads the device read-only, and only sgdisk -p is unavailable.
  • Unverified hypothesis, not a conclusion: historical evidence shows an Android ramoops region at 0xa7000000-0xa7400000, so a kernel log kept across a warm reboot could be read from Android /sys/fs/pstore without the storage driver; the current FD window 0xa7100000-0xa7400000 lies inside that region and the memory ownership gate is open, and there is no evidence whether a reboot clears DRAM.
  • Writing the reserved log region is a device write and needs explicit user authorisation in the round that performs it.

Next action

  • Re-run the record and diff checks after the commit, and add host or device evidence only when an actual build or device run produces it.
  • User decision (2026-11-07): after the kernel checkpoint, capture kernel logs persistently instead of repeated screen colour-band or hex recordings, following docs/zh/porting/log-capture.md sections 4-6: reserve a log region inside userdata, connect the kernel writer to the storage driver's ranged read/write/sync path, and read it back with a minimal Linux image or a recovery+adb reader, with the writer, reader and host decoder sharing one block unit and byte range.
  • Until the route B ShengUfs storage driver works, keep the firmware Tools -> Advanced Logger QR export plus serial/screen checkpoints; the [userdata] first_lba/last_lba/lba_bytes geometry is already recorded in board.toml, so the remaining step is to choose and reserve the log region inside it; reference the official UniXNU-sheng package/BootLog.c, BootReset.c and package/linux-log/ for structure only, because its PMIC addresses and watchdog calls are sheng board values that nuwa must replace with its own evidence. The Advanced Logger producer is now wired (uefi-nuwa e494c01b2f679439fdaeaf2084b693c7220543f5), so the next check is the runtime Tools -> Advanced Logger page on hardware.
  • Next host phase (decider CONTINUE implemented, 2026-11-11, self-checked): the explicit pmap-ret-continue-20261111-r1 profile takes the independently frozen renderer baseline 7fd3d933aae66d8e34a0d3ebfa142137978976d7521e04cef20321109750b6f4 (55386112 B), restores the normal PMAP RET and the observer self-loop from the verified diagnostic 6671eafeab4e1ad73ba73f8c85728f6f7f2fc78a7db9398721b911e1b613f269 by exactly two inverse 8-byte words, and sets the effective arguments to -v serial=3 only (the old sheng-platform-stage=5 is dropped because ArmDTPlatform.cpp:200-209 resets the stage after registerService returns). New ignored host FAT 3e98ba12f13df9a776af5de85e254ff3b95fc94a7bc38eb53edd1aca774521f1, ADT a296706d390447319c11a9db7b691f9d4ecc3276bac6424ad9e04598c16aab7a (13256 B), schema2 record fb326d3441b0d9a552463c84d0b4d240e68937c54342097acbb28f78992c97a5; 5 increment tests/CLI/new profile-aware audit exit 0. The independent verifier then passed its 5 tests/CLI/FAT/layout with package FAT 9ff5c2720a1af341b66e8417a3308fc0a3be9e561c3406e7db157d05e523d04c, ADT 885094de4597fb2d1734ba4092c805968097d021c8f482653b238436586d27a4 and record a7ca5cc8a15353a9dee8882a1f89e0633a44ff950514ebe6ad58c4567ac20b57. The continuation outer vendor_boot is now frozen host-only at out/owner-continuation-vendor-boot-20261111-final/vendor_boot.img, 100663296 B sha256 4a52ddbc2cc06401f76676f480eb1b0e4b1dfafce0a8d07b453dd53967a05618, with the actual vendor ramdisk exactly the verifier FAT and DTB/bootconfig/vbmeta/footer preserved per the reviewed allow-list; 4 increment tests and full-chain readback exit 0. Still no device authorization/executor, no runtime mapping or RET execution claim, authentication NOT REACHED. Interpretation contract: the goal is the first natural panic or a visible native milestone; MAC Framework successfully initialized only proves the mac_policy_initbsd printf before callbacks; unseen logs are not evidence; menu stop/reboot is not automatically PMAP/watchdog/stage5; no output is ambiguous, not 'not reached'; at most one bounded BootArgs.Video/PE_state.video/PE_create_console/console_write check if the real follow-up shows no output. Read-only upstream survey: the SM8550 common checkout is ahead 6 and XnuPrelink ahead 5 of the effective pins; the pins are unchanged and the impact is undetermined, so do not record them as unrelated.

Evidence

cellular

planned

Components: QcomBaseband.

Hardware and dependencies

  • Modem firmware and EFS/NV images extracted from the device.
  • QRTR/QMI channel and IPA/QMAP data path evidence from the vendor DTS.

Configuration

  • board.toml and DTS entries for the modem, plus the kernel command line switches selected for it.

Procedure

  • Follow the baseband guide after storage, display and input are device-verified.

Development findings

  • 2026-11-15 no cellular work: this commit only narrows the area paths from package/** to cellular-specific files, because the same commit lands package/NuwaPlatformPower.cpp and package/diagnostics/* which belong to the power and diagnostics areas. No modem, EFS/NV or QMI evidence exists.
  • No modem evidence recorded yet; SM8550 shared support does not carry this board's modem firmware. package/README.md added by the format-2 migration.
  • proprietary-files.txt now lists only the Binaries/nuwa DXE set from the public Device-Binaries repository; no modem firmware or EFS/NV input is recorded in it yet.

Completion criteria

  • SIM identity, network registration and a data session are observed on the device.

Current limits

  • Planned; outside the first boot milestone.

Next action

  • Extract modem firmware and EFS inputs, then record QRTR/QMI channels from the vendor DTS.

Evidence

Collect evidence using the next action.