feat!: remove the build.config path and gate upgrades on the record - #21

Refs #13 — step 3 of three, the last of the keystone. Targets release/1.0.0.

Steps 1 and 2 are merged: Castle materialises the target's configuration in a :peer
running the target's own code, and Forecastle no longer intercepts configuration at all,
so no release it assembles has build.config any more. This is the cleanup that promised,
plus two changes Forecastle's work specified along the way.

Removed: the build.config path

Castle.Commands.generate/1, expand/1, write_sys_config/2, the build.config branch
of materialise/2, and the public Castle.generate/1. Dead with them: the config
provider stub and its tests.

Nothing called it — Forecastle's only remaining reference is a negative assertion,
refute env_sh =~ "Castle.generate", a test that the call is gone.

materialise/2 is now one path. The surviving case is "there is no release here", and it
uses File.ls rather than File.dir? deliberately: the peer's own refusals name a
missing file, which is right for a release that was unpacked and then damaged and wrong
for one never unpacked — and an empty directory belongs with the absent one.

Changed: make_releases no longer depends on the working directory

release_handler resolves relative paths against :code.root_dir(), not the cwd —
consult/2 and do_write_release/3 both go through root_dir_relative_path/1. So the
only cwd-sensitive thing here was ever its own File.exists? guard, and callers no longer
need to change directory to use it. That matters because the working directory is
VM-global: doing this over rpc on a live node meant any application performing relative
file I/O during the call resolved under the release root.

It must stay create_RELEASES/3, never /4 with the root supplied. /3 is
create_RELEASES("", RelDir, RelFile, LibDirs), and check_rel_data/4 branches on an
empty Root to store lib dirs relative — OTP's comment says this is "to make it easy to
create a relocatable RELEASES file". Verified by experiment rather than argument: /3
with an absolute directory wrote lib/kernel-10.5; /4 with the root wrote
/var/folders/…/lib/kernel-10.5. Passing the root would bake absolute lib paths in and
silently make the release non-relocatable, which no end-state test would catch. The code
comment and a unit test both say so.

The directory is an argument rather than self-derived, matching this repo's existing
convention — under mix test, :code.root_dir() is the OTP installation, which has a
releases/RELEASES, so a self-derived directory short-circuits every test to {:ok, []}
and makes the error paths untestable.

Added: Castle.upgradable/0

A system whose releases/RELEASES was missing when it booted runs from a record OTP
synthesised from the boot script — which names no applications. Upgrading from that
state reports success and leaves any application whose version changed, but whose code
the relup does not explicitly load, running its old code.

Confirmed in sasl-4.3: init/1 falls back to a #release{} whose libs is [];
which_releases reports mk_lib_name(Libs) and mk_lib_name([]) -> []. And the reason
it is silent rather than loud: get_new_libs([], _) -> [], and point_of_no_return calls
code:replace_path/2 only over newlibs.

So the discriminator is exact rather than heuristic — a synthesised record reports an
empty application list, and a real one cannot, because kernel and stdlib are always
present.

upgradable/0 refuses in that state and names the remedy. Note what it must say and why:
the file cannot be repaired from a running node, because release_handler reads it once,
in init/1. Creating it now changes nothing about the record this node works from. The
system has to be restarted — and the release creates the file before it starts, so a
restart is all that is needed.

This replaces a shell-side approximation in bin/castle, which tested whether the file
existed. That passes in exactly the case the gate exists to catch: a file that appeared
after the boot which looked for it.

Docs corrected rather than appended to

AGENTS.md: the two-shape dispatch is no longer described as load-bearing (it records
that the second path existed and must not come back); the write_sys_config umask note
went with the code it described; the concurrent-boot sys.config race is deleted,
because nothing of Castle's runs at boot now; the cold-boot limitation is re-scoped from
"unreachable" to "reachable, untested, and Forecastle's to test"; and a stale claim that
the pinned Forecastle "never calls this" is gone — it does.

RELEASE.md: generate/1 removed from the entries that described it, a Removed entry
for the path, a Changed entry for the cwd independence, and an Added entry for the
gate. README left alone; its staleness is tracked under #9.

Verification

mix precommit green — 86 tests. Compiled and tested against Elixir 1.20.0/OTP 28
and 1.18.3/OTP 27 as well: clean, 86 passing on both.

Cross-repo, which is the only thing that exercises the peer path end to end: Forecastle's
suite with its fixture pointed at this branch, --include e2e, 170 tests, 0 failures.
The installed version carries CASTLE_MATERIALISED=true and a sys.config.pristine, with
Mix's own header preserved.