A transient runner DNS glitch could fail just the upload step for one
platform while its build and checksum succeeded, and the run still finished
green with that platform quietly missing from the release. The "attach to
the release" step now compares the platforms actually present in dist/
against the full expected set and emits a ::warning:: naming whatever is
missing, pointing at release-all-missing.yml to build the gap - instead of
requiring someone to notice by comparing sixteen build jobs' logs by hand.

The released PowerPC binary reported its version but aborted while
initializing V8. Real target snapshot tools are now built under QEMU, as for
s390x, and both platforms' artifacts are rejected unless JavaScript,
separate V8 contexts, crypto and compression actually execute - the gate
reproduces the released failure and passes with official same-version Node.

Release All expands from seventeen to forty-two OS/CPU targets after
compiling current upstream master with Go 1.27 across Go's native
command-line platforms, adding AIX, DragonFly BSD, NetBSD, OpenBSD, all
FreeBSD CPUs, and Linux MIPS and big-endian PowerPC. Illumos, Solaris and
Plan 9 are excluded because current upstream source does not compile there.

The CGO-free upstream tools compile for Android arm64, expanding the
canonical registry to forty-three targets and 344 possible binaries. Other
Android architectures still require external CGO linking; iOS and
WebAssembly do not produce equivalent standalone command-line programs.

Both full and missing-only workflows now allow three hours for the expanded
forty-two target build instead of the former seventeen-target one-hour
limit, so adding platforms cannot create a predictably cancelled release.

Upstream mongodb/mongo-tools was checked for an analytics client, a
phone-home reporter, or a telemetry/DO_NOT_TRACK flag of its own, the same
way wekan/mongosh-patches and this fork's FerretDB were checked before
their telemetry was patched out - there is none. tests/no-telemetry-upstream.sh
re-checks this against the current upstream ref so a future release that
adds real telemetry is caught here instead of silently missed.

The bundled analytics sink is now unconditionally a no-op, so no telemetry
HTTP request is ever made regardless of the configured endpoint, and the
native machine-id lookup that only existed to key telemetry throttle state
is dropped. The startup banner says this fork does not collect or send
anything. The same patch fixes mongosh running inside ghcr.io/wekan/ferretdb
as its default non-root user, which has no /etc/passwd entry: config/log/
history storage now falls back to a writable directory under the OS temp
dir when the home directory is not writable, instead of failing with
EACCES ... mkdir '/nonexistent' and "Could not open history file" on
every session.