Add ramdisk-specific kernel snapshot and build logic so the installer ramdisk can boot with a conservative kernel while the restore target keeps the fully JB-patched kernel. Changes:
- research/patch_comparison_all_variants.md: document the Ramdisk Kernel Split and intent.
- scripts/fw_patch_jb.py: snapshot the base/dev-patched kernel before applying JB extensions (new helper and constants).
- scripts/ramdisk_build.py: build krnl.ramdisk.img4 from the snapshot and krnl.img4 from the restore kernel when a snapshot exists; factor kernel IMG4 creation into build_kernel_img4.
- scripts/ramdisk_send.sh: prefer krnl.ramdisk.img4 when present, falling back to krnl.img4; fail early if no kernel image found.
This improves /dev/disk1s1 remount reliability during CFW/install by keeping the restore kernel JB-patched but booting the installer ramdisk with a more conservative kernel variant.
Three bugs caused NOT_BOOT (timeout):
1. Wrong sysent table base: first _nosys match is entry 428, not entry 0.
Entry 0 is the indirect syscall handler. Fixed with backward scan.
2. Raw VA written to chained fixup pointer slot: struct.pack("<Q", cave_va)
corrupts the fixup chain from sysent[439] onward. Fixed with proper
auth rebase encoding (_encode_chained_auth_ptr).
3. Missing PAC parameters: dispatch uses BLRAA X8, X17 with X17=0xBCAD.
Chained fixup must encode diversity=0xBCAD, key=IA, addrDiv=0.
Chain 'next' field preserved from original entry.
Boot-tested OK via testing ramdisk.
Root cause: find_string("vnode_getattr") matched "%s: vnode_getattr: %d"
format string inside an AppleImage4 function. The old code then took that
function as vnode_getattr itself, causing BL to call into AppleImage4 with
wrong arguments → PAC failure on indirect branch at a2+48.
Fix: _find_vnode_getattr_via_string() now scans backward from the string
ref for a BL instruction and extracts its target — the real vnode_getattr
(sub_FFFFFE0007CCD1B4 at foff 0xCC91B4).
Bisection confirmed: variants A (stack frame) and B (+ tpidr_el1) boot OK,
variant C (+ BL vnode_getattr) panics with old resolution, boots OK with fix.
Boot-tested: full C23 patch with corrected vnode_getattr — BOOTS OK.
Root cause: backward branch search found PAC validation B.EQ (target fell
within error region range) instead of the kernel_map guard B.NE. New approach
walks backward from panic string ADRP to find CMP+B.cond where the branch
goes forward past the panic, then makes it unconditional. Also enables B6
and B8 in find_all(). Boot-tested: PASS.
Old approach patched vnode-type assertion BLs (CBZ→panic). New approach
scans backward from function end for the shared MOV W0,#1 kill return
before the LDP x29,x30 epilogue and changes it to MOV W0,#0. Single
instruction converts all 5+ kill paths to success. Boot-tested OK.
Root cause: "most-called BL target" heuristic in _proc_info picked copyio
(4 calls, 0x28C bytes) over the real _proc_security_policy (2 calls,
0x134 bytes). Lowered size filter threshold from 0x300 to 0x200 to
correctly exclude utility functions like copyio. Boot-tested: PASS.
The original selector24 patches (NOP ldr + NOP bl) broke the hash flags
extraction, causing the consistency check to fail. The second attempt
(return 0x130A1) also failed because the return code semantics were
inverted — byte 1 != 0 means FAIL, not success.
Correct approach: insert `mov w0, #0xa1; b <epilogue>` after the prologue.
0xA1 has byte 1 = 0 which the caller checks via `tst w0, #0xff00` as PASS.
Update AGENTS.md
move selector24 bypass from txm_jb.py to txm_dev.py, delete TXMJBPatcher
Selector24 CS validation bypass now applies to both dev and JB variants
via txm_dev.py. The separate txm_jb.py patcher is removed since it had
no other patches. Dev boot chain: 47→49 patches.
Create txm_fullchain_analysis.md
