Rejustifies the duration subset rule

[johnnyt]

Why

sui-cw0 (mirrors px-69c) was filed while predicator's evaluated durations
had an expression-dependent key set: seven units normally, eight when the
expression mentioned ms, against a declared type of eight. px-69c closed on
the eight-key branch (predicator PR #176, shipped in predicator 9.0.0), so the
justification this repo wrote against the seven-key output is now false, even
though the code it justifies is still right.

What

StatifierUI.Shape.duration?/1 still accepts any non-empty atom-keyed subset
of the eight units. The @doc no longer explains that by citing a parser bug;
it explains it as the viewer's own tolerance - statifier-ui renders a value
stream it did not produce (a hand-written fixture, a decoded wire message, a
datamodel from an older engine or another interpreter), and a duration short a
unit should read as a duration rather than collapse into a seven-field map.

The pinned test "durations from real predicator expressions" was passing
incidentally and now pins the rule explicitly. See the sabotage evidence below.

Contract re-read, with citations

Resolved dependency: mix.lock:27 pins predicator at 9.0.0 (>= 9.0.0, as
required).

• deps/predicator/lib/predicator/types.ex:37-46 - @type duration declares
  all eight units, milliseconds: non_neg_integer() among them, and the
  moduledoc at types.ex:24-29 states "Every unit the expression did not name
  is present and 0, so the map is always this wide".

• deps/predicator/lib/predicator/duration.ex:36-48 - new/1 builds all eight
  via Keyword.get(opts, :milliseconds, 0).

• Live probe against the resolved dependency, all eight keys every time:

  "3d"       -> 8 keys
  "2w"       -> 8 keys
  "1h30m"    -> 8 keys
  "3d8h"     -> 8 keys
  "500ms"    -> 8 keys
  "1s500ms"  -> 8 keys
  

Sabotage evidence

Temporarily tightening duration?/1 from MapSet.subset?/2 to
MapSet.equal?/2 (require exactly eight keys), then reverting:

• Before this branch's test change: 3 failures - the module doctest, the
  seven-key test, the partial-duration test. "durations from real predicator expressions" stayed green. It was passing incidentally, which is the
  outcome this bead was filed to prevent.
• After this branch's test change: 4 failures - the same three plus
  "durations from real predicator expressions", failing on
  Shape.infer(Map.delete(value, :milliseconds)) == :duration.

The rule is reverted to MapSet.subset?/2; the sabotage was an experiment, not
a change. The test now asserts both halves: that real evaluator output carries
all eight keys (so a regression upstream fails at the dependency), and that
dropping a unit from that real output still infers as a duration (the tolerance
the subset rule exists for).

Value.encode/1 is correct as it stands - unchanged here

lib/statifier_ui/value.ex:191-196 maps over all eight @duration_units with
Map.get(duration, unit, 0), so encode/1 fills an absent :milliseconds
with 0. Under the eight-key contract that writes a unit the value domain
declares (types.ex:45) and that Duration.new/1 itself produces - it is
canonicalization, not invention. The bead's encode-side worry applied only to
the seven-key branch, which predicator did not take. No change was needed and
none was made; the file is untouched on this branch.

Residual, deliberately not fixed here: value.ex:24-29's moduledoc carries
the same stale parenthetical ("predicator's parser emits seven, omitting
:milliseconds"). It is prose only and outside this branch's declared file
set, so it is reported rather than edited. Worth a follow-up bead.

Sibling bead

st-4epq in statifier-ex (also mirrors: px-69c) is CLOSED - "Acceptance
criteria satisfied via st-793d (predicator 9.0 consumption bump, merged PR
#175); no engine-side work remained". Its notes record a seven-versus-eight
audit of that repo finding no key-set assumptions. Nothing is owed there.

Notes

• Full mix quality green on the rebased HEAD (attested, not scoped): format,
  compile, dependencies, doctor, credo, 726 tests / 92.9% coverage, dialyzer.
  Gettext and Sobelow are the two permanent not-applicable skips.
• No changelog fragment: doc and test only, no public API or observable
  behavior change, per changelog.d/README.md.
• Rebased onto main at 4ad5a9d; no commits replayed, no conflicts.
• Mirrored with px-69c - not closed here. No Closes line by design:
  this half stays open for the operator.

Refs: sui-cw0