feat(tooling): derive each policy's input fields from its AST

[kmadan]

The problem

"What does this policy need me to supply?" was answered by a hand-written
# RequiredMetrics: comment block at the top of each policy. That block had
drifted: 22 of 98 policies read input and declared nothing at all.

Consumers got an empty list and no way to know it was wrong. Concretely, in the
playground two EU AI Act policies (transparency, risk_management) rendered
no questions at all and passed on values the visitor could neither see nor
change. This is the same class of drift that the coverage figures had before
--check existed, in a different field.

The fix

scripts/extract-input-fields.sh derives the fields from the policy's AST via
opa parse --format json. Three access forms are recognised:

input.system.high_risk                          # a plain ref
object.get(input, ["system", "sources"], [])    # a path passed as an argument
object.get(input.params, "threshold", 0.8)      # a ref plus a literal key

The second form is used 284 times in this library — it is how a field is
read without the whole rule going undefined when the field is absent. A regex or
a naive ref walk reports almost nothing for those policies.

It also infers the kind of each field from the literal the policy measures it
against: the default handed to object.get, or the constant on the other side of
a comparison. input.system.x == true is a yes/no question; object.get(input, ["logs","months"], 0) >= n is a number. That is firmer evidence than reading a
sample document, which can omit the field entirely.

Why the comments stay

Neither source is complete alone, so consumers union them.

The AST cannot recover a field name that is computed:

some label, field in criteria
object.get(input, ["datasets", field], false) != true

Here the names live in the criteria map, and the comment is the only static
record of them. The AST contributes a bare datasets for this call, so:

• prefixes are pruned against the fuller list, since datasets is an object
  rather than a question and is answered by any leaf beneath it; and
• the kind found on the prefix is inheritable by the leaves — false as the
  object.get default makes every datasets.* leaf a yes/no. This resolves 51
  of the 53 EU AI Act fields that have no evidence of their own.

Result

Framework	Before	After
EU AI Act	142 declared, 2 policies with no fields	151 declared, 0 with no fields, 123 typed
UK	31 declared	33 declared, all typed

Two fields remain untyped: logs.role (covered by the sample) and
logs.sectoral_minimum_months, whose object.get default is a variable rather
than a literal. Consumers fall back to a text input, which is honest.

input.params.* thresholds move to a separate parameter_fields list. They tune
how strict a check is and each has a default in the Rego; listing "toxicity
threshold" among the obligations would invite someone to answer their way to a
pass.

Verification

• opa check, regal lint (196 files, no violations), 806 tests,
  generate-coverage.sh --check, check-version-refs.sh
• build-playground.sh --verify: all sample verdicts unchanged. That is the
  main safety property — this adds knowledge about the policies without
  altering what any of them decides.