The package installs and launches, and the eboot shut the console down

The install and launch path is solved. What remained was the console's own state and one build-time default, and the second is the more useful surprise.

Four things had to be true before launch OBSC00001 got past the launcher, and each was found by moving the error code forward one step at a time (0x80020000 | errno, so ...0002 is ENOENT and ...0016 is EINVAL; the sceSystemServiceLaunchApp: Resource temporarily unavailable string mislabels every one of them):

• The nested pfs had to mount. Four selfish fixes, each measured against real packages: the outer inodes carry 0x0C (this crate wrote 0x10), the inner ones 0x10 (it wrote nothing), the inner superblock's embedded inode had to be filled at all, and pfs_image.dat needed the compressed bit with size_compressed holding the inner length.
• kstuff has to be loaded or the fake licence is refused - sceRifManagerStoreRif 0x80870003 into an AppPromoter error into preLaunchCheck 0x80a40086. It is first in /data/pldmgr/autoload.txt, but a plugged-in USB with its own payloads takes precedence (SCAN_USB_PAYLOADS=1) and that chain omits it.
• No other app may hold the display. With everything else right the launcher gets through login, budget and directory checks and then fails checkExistingApp: 0x80940010 while the resource dump shows another app RUNNING with vcn[20] scanin[2K]. One big app at a time.

The surprise: make pkg builds the suite that CLAUDE.md warns about

On the one launch that the launcher accepted, the console shut down and showed CE-108262-9 on reboot - "an error occurred while reading the system software or application data". The package was not the cause: it is self-consistent (13/13 entry digests, both manifest digests, every size field agreeing) and its entry layout is byte-identical to a real package's.

The cause is a default. make pkg builds with EXCLUDE empty and SERVE unset, so the eboot takes the #else branch in start.c: obs_run_all() over all 146 checks with no exclusions, then obs_screen_present(). This file already says building without the exclusion list "has twice handed a later step a module that walked into a known crash", and this is the third - the first time in a foreground app that owns the display, where the cost is the whole console rather than one process.

There is no excludes.txt on this machine to apply, and the sweep that generates one wants a loader that survives the run - which is the thing being established. So the eboot is now built SERVE=1, which is the mechanism start.c describes for exactly this: listen, run nothing, and let a driver ask for the suite over the socket, so "a crash during it costs one session and a reconnect rather than the whole endpoint".

start.c also announces itself on the kernel log now (sceKernelDebugOutText, guarded like every platform call, import registered in imports.c) before the entry, the net check, the HUD and the serve loop. Principle 1 applied to the boot sequence: the report channel does not survive a fault and the log does. None of these breadcrumbs has printed yet - the eboot has not run since they were added, so they are the instrument for the next launch and not evidence of anything.

The better instrument was already here and was overlooked for three crashes: sink.c writes the report unbuffered, one syscall per record, to /data/obscene-report.txt and three fallbacks. That file outlives the crash and the reboot, so whether the eboot ever ran is answerable after the fact without watching the screen - scripts/oops-recover.sh reads it and looks for a crash dump.