Adds a payload-schema drift test for the doc

[johnnyt]

Why

docs/wire-format.md is the format's normative home: when it and an
implementation disagree, the document is what conformance means. But only its
type index was mechanically guarded (wire_format_spec_test.exs). Nothing
tied its per-table field, type, and presence schemas to what
StatifierUI.Trace.Manifest actually builds, so drift in those tables could
only be found by reading them. The sui-o5c verify walk found two such prose
defects by hand, which is what filed this bead.

What

A sibling test, test/statifier_ui/trace/wire_format_payload_schema_test.exs,
parses the six markdown tables of the session.start section - the top-level
payload table, states, transitions, contents, data, and the location
object - and checks each against three built messages: a chart written to
reach every conditional field, a bare one written to reach none, and the
first of those projected (so the projection header has somewhere to be
observed present).

It parses the document rather than restating it, the way the type-index test
already does, so the document stays the single place a field is written down.
Three properties are checked per table:

• a field the producer emits with no doc row fails;
• a doc row nothing emits fails;
• the presence column is enforced both ways - a row marked always must
  be present on every sample row, and a row marked present only when ... or
  omitted ... must be observed both present and absent, so a conditional
  row that has quietly become unconditional fails rather than passing
  vacuously.

A fourth guard test pins the set of table names the parser resolved. Without
it, a restructured section that parsed to nothing would make every other test
here iterate an empty schema and pass.

The one disagreement it found

data.value_location was documented as present only when the compiler
recorded a span for the element's value. That is stale: every arm of
Statifier.Compiler's build_data_value/2 returns a Location.t() - the
written value's span for an expr- or src-written element, the <data>
element's own span otherwise - and the document's own prose immediately below
the table already described exactly that fallback. The table cell contradicted
its own section.

Resolved on the document side, which owns this contract: the presence cell now
reads always, with the fallback named. No producer was changed - the code
was right and the table was stale.

Notes

• Full mix quality green and attested on the rebased HEAD (909 tests, 92.9%
  coverage). The two ○ lines are the repo's permanent Gettext and Sobelow
  skips.
• The test was verified to bite rather than merely pass, by six mutations run
  and reverted: an extra producer field with no doc row, a phantom doc row, an
  always cell flipped to conditional (on data.d_index and on
  states.children), a conditional cell flipped to always
  (transitions.cond_location), and a renamed table label that makes the
  parser find nothing. All six failed the suite.
• Scope note: this covers the session.start payload tables only. The nine
  trace.* and ten effect.* schemas are the same shape of gap and would be
  a follow-up bead, not a silent extension of this one.
• A changelog fragment is included: the presence correction changes what the
  published wire-format contract promises a consumer, even though no emitted
  bytes moved.

Closes sui-o4e