The Boost.Context 1.91.0 sync (phpGH-22584) pulled in boostorg/context#325,
"fix cortex-m0 support to aapcs_elf_gas target", which adds a fallback
branch to make_fcontext/jump_fcontext for the Thumb-1 encoding limits of
ARMv6-M. The branch is correct for Cortex-M0 but is gated on
__ARM_ARCH >= 7, which is also false for classic ARMv6 in ARM state
(Alpine armhf, Raspberry Pi 1/Zero) -- targets that can encode every
instruction in the >= 7 branch and need none of the workaround.

Two consequences. make_fcontext took `ldr a2, =finish', materialising the
absolute address of finish through a literal pool in .text and emitting
R_ARM_ABS32 against .text. In a PIE link that becomes an R_ARM_RELATIVE
dynamic relocation inside the read-only text segment, tagging the binary
DT_TEXTREL. musl supports DT_TEXTREL only for DSOs it maps itself; the
main executable is kernel-mapped and handled by kernel_mapped_dso(),
which never inspects it, so ldso writes to a read-only page and dies with
SIGSEGV in do_relocs() before main(). Every SAPI, every invocation, no
output. glibc masks this by mprotecting the segment writable, which is
why only musl builds noticed.

Separately, jump_fcontext -- the hot path on every fiber switch -- grew
from 13 to 43 instructions on all ARMv6 builds, glibc included.

Gate on __ARM_ARCH_6M__ instead, exactly the target php#325 names. The
result emits byte-identical .text to 8.6.0alpha1 on every ARM target
except Cortex-M0, where it emits byte-identical .text to current master:
each target gets back the code it had before the regression, and M0 keeps
what php#325 added. armv7 and arm64 are unaffected either way.

Zend/asm is vendored and verified by the Verify Bundled Files workflow,
so carry the same diff in .github/scripts/download-bundled and re-apply
it from the sync script, as uriparser already does.