fix: run Claude Code where Android refuses its syscall

[rmyndharis]

The Claude Code sign-in panel reported Claude Code process terminated by signal SIGSYS on Android 13 and 14, and a terminal printed Bad system call. Neither
says what happened, and both read as if VSCodroid broke something.

Root cause

An Android app may make only the system calls bionic exposes in SYSCALLS.TXT,
and epoll_pwait2 (441 on aarch64) appears there in android15: it is absent
from android13-release and android14-release. The CLI's runtime calls it as
soon as its event loop starts, so the process was stopped there. That is why
claude --version printed and exited happily while anything that ran the loop
died.

Established with a ptrace tracer run inside the app (it has to live in
nativeLibraryDir, since SELinux refuses execve under filesDir):

TRACER: SIGSYS in tid 21248 -- si_syscall=441 x8=441

Each alternative was measured and ruled out, not argued away: the extension
build from four months earlier fails identically; both emulators run a
byte-identical APK; DNS and HTTPS resolve on the failing device; targetSdk has
been 36 since 1.1.0; and the failure reproduces with the CLI sandbox both on and
off.

The fix

Android refuses with SECCOMP_RET_TRAP rather than a kill, so the refusal
arrives as a catchable SIGSYS: the syscall does not run, and a handler that
returns resumes the thread with whatever it left in x0. That is enough to
answer the call.

• scripts/seccomp-shim.c builds libseccomp-shim.so, preloaded into the CLI.
  It emulates epoll_pwait2 with epoll_pwait, rounding the timespec up to the
  next millisecond so a sub-millisecond wait does not become a spin. Any other
  refused call is left to die as the platform intended: the handler restores
  SIG_DFL and returns.
• It also interposes sigaction. The CLI installs a SIGSYS handler of its own
  that re-raises through kill(); without the interposition the emulation
  answers three calls and the fourth is fatal. Every other signal is passed
  through with the layout translated in both directions, since musl's
  struct sigaction is not the kernel's.
• scripts/claude-launch.c builds libclaude-launch.so, which is what
  claudeCode.claudeProcessWrapper now names. It appends the shim to
  LD_PRELOAD and execs musl's loader. Setting LD_PRELOAD globally would put
  the shim in front of Bionic children too, which would read its musl-shaped
  sigaction translation and corrupt their own.
• The shim links with --pack-dyn-relocs=none. Targeting API 30 or newer, lld
  emits relative relocations in Android's packed form (DT_ANDROID_RELR), which
  musl 1.2.5 does not implement: the constructor pointer in .init_array keeps
  its link-time address and the CLI dies with SIGSEGV before running a line. The
  build fails on either that or a NEEDED entry, which would put a second libc
  in a musl process.
• Both files are requiredJniLibs, so an APK cannot ship without them, and
  build-all.sh, build.yml and release.yml build them alongside the other
  native pieces.

docs/USER_GUIDE.md now says the platform limit is handled and what to report if
a future CLI reaches for a call the shim does not cover. The claude wrapper
keeps its 159 check for the same case.

Verification

Measured on an API 33 emulator with an API 37 control, through the shipped
launcher rather than a hand-built one:

	before	after
claude --version	killed, signal 31	2.1.240 (Claude Code), status 0
stream-json session	killed, signal 31	full init JSON, then its own auth prompt
sign-in panel	terminated by signal SIGSYS	signs in
SIGSYS in the log	on every attempt	none

API 37 is unchanged: the syscall is allowed there, so the handler never fires.

2227 JVM tests, 0 failures; lint clean; the punctuation, document-date,
build-step, workflow-step, translatable-string, bundled-extension, binary-list,
checklist and instrumented-inventory gates and the nine JavaScript self-checks
all exit 0.

Sources: bionic SYSCALLS.TXT on the android13, android14 and android15
branches; asm-generic/unistd.h for the syscall number.