Guards the package files: list against ADR-0009

[johnnyt]

Closes sui-2ke.

Provenance

Campaign 027, statifier-ui lane. Cut from origin/main at 7cbf911; one
commit, no rebase needed (origin/main had not moved at push time), so no
sibling conflict arose. A sibling sui worker may be live on lib/ and
docs/ for the wire-format decisions; this branch touches neither.

Files touched: test/packaging_test.exs (new, 58 lines). Nothing
else - mix.exs is deliberately unchanged, see below.

What this is

ADR-0009 ("JavaScript ships as source") decides that this package's JS is
delivered as source under assets/ and that "assets/ becomes public
API". mix.exs's package() files: list has never named it. Hex
packages exactly what files: names and reports nothing about what it
leaves out, so the first commit to ship JavaScript would have shipped it
into a tarball that silently omitted assets/, with the first symptom
being a host that cannot resolve the import.

Why the files: entry is not in this PR

The bead's original fix was to add assets to files:. Probed on this
branch: hex refuses to build a package whose files: names a directory
that is not on disk.

$ mix hex.build
Building statifier_ui 0.2.0
...
** (Mix) Stopping package build due to errors.
Missing files: assets

(hex 2.5.0, Elixir 1.18.3, OTP 27.3. mix hex.publish shares that code
path.) assets/ does not exist yet, so adding the entry today would break
the next release in order to prevent a later one. Committing a placeholder
assets/ was the other option and is barred: ADR-0009 names no layout for
this package's own assets tree - every assets/package.json in the record
is the host's file - so a placeholder would invent the first slice of a
declared-public-API surface, which is a contract call rather than a
packaging fix.

The bead was re-scoped to the guard only. The first JS ship (sui-wqr, or
the ADR-0008 elkjs renderer) adds the files: entry alongside its first
real asset, and this test is what makes forgetting it impossible.

The guard

test/packaging_test.exs asserts that every top-level directory an
accepted ADR declares published appears in files: once it exists on
disk. The declared set is a tracked module attribute citing its record,
not something inferred from the filesystem, and a second test keeps that
list from being quietly emptied into a vacuous pass.

Verified in all three states:

assets/ on disk	in files:	result
no	no	green (today's state)
yes	no	red, message names the directory and ADR-0009
yes	yes	green (the state the first JS ship lands in)

The red case was produced by creating assets/index.js, running the test,
and removing it again; the green-green case by adding assets to files:
at the same time and reverting both. Neither probe is in the commit.

Changelog

No fragment. changelog.d/README.md excludes "test harness or internal
test-support changes", and nobody calling the public API can tell the
difference - the packaged tarball is byte-identical to before this PR.

ADR

No ADR touch. The record already says everything the guard relies on
("assets/ becomes public API"); the test's own header comment carries the
hex-refuses-a-missing-directory finding, which is a toolchain fact rather
than a decision.

Gate

Full mix quality green on this HEAD, attested (mix quality.verify,
data.attested: true): 880 of 880 tests, 92.8% coverage, dialyzer clean,
credo strict clean, doctor passed. The two ○ lines are the permanent
Gettext and Sobelow skips declared in gate.not_applicable_skips.