[flagd] The two resolvers disagree on numeric coercion: in-process returns TYPE_MISMATCH, RPC returns 0

Confronta i percorsi del resolver in-process e RPC del provider usando il caso get_integer_details descritto nell’issue, iniziando dai risultati esistenti della conformance-suite e dalla voce relativa alla deviazione nota. Verifica il flagd coercion ADR e il comportamento del resolver rispetto a flagd-testbed:v3.8.0. Il lavoro è completato quando il contratto previsto è stato deciso e i risultati dei due trasporti sono allineati, oppure quando la deviazione rimanente è documentata esplicitamente.

The two resolvers of this provider answer the same question differently, and I would like to know which one is intended before assuming either is wrong.

What I measured

float-flag resolves to 0.5. Requesting it through get_integer_details with a code default of 1:

Both against flagd-testbed:v3.8.0, same flag, same call, same run. The widening direction agrees:
integer-flag (10) requested as a float returns 10.0 from both.

So in-process rejects a lossy narrowing and RPC performs it silently.

Why I am asking rather than filing it as a bug

The rule this would be measured against is flagd's own, not OpenFeature's. OpenFeature has a
single numeric type — number is "a numeric value of unspecified type or size" — and no numbered
requirement says what a provider must do when a value does not fit the accessor it was asked
through; that is an open question of the provider contract
(open-feature/spec#430). flagd accepted a coercion
ADR for itself — coerce when lossless, TYPE_MISMATCH when information would be lost — tracked in
open-feature/flagd#1996.

Against that rule the in-process resolver is already correct and the RPC one is not. But that is
flagd's rule to interpret, which is why this is a question.

What does not depend on which answer is right: two transports of one provider disagreeing is a
vendor-neutrality problem on its own terms. An application that switches resolver= — a
configuration change, not a code change — gets a different value and a different error code for the
same flag and the same call. That is the property the conformance suite exists to check, occurring
inside a single provider.

Calibration from the other languages

Worth knowing before deciding, because it suggests where the fix belongs:

• The Java and Go flagd providers narrow 0.5 to 0 with no error code in both resolvers
  — consistent with each other, both matching this repository's RPC behaviour.
• This provider is the only one of the four whose two resolvers differ, and the only one where any
  resolver implements the ADR's lossy half.

So the in-process path here appears to be ahead of the other implementations rather than behind them.

Questions

1. Is the in-process behaviour the intended one, with RPC yet to catch up?
2. If so, is the gap in this provider's RPC path, or is it that the evaluation happens server-side
   there and the answer has to come from flagd itself?
3. Is there a reason the two paths should be permitted to differ that I am missing?

Context

Found while building the cross-language provider conformance suite proposed in
open-feature/spec#417. Both resolvers now declare
the @numeric-coercion capability and the results record the difference, with a known-deviation
entry against the RPC suite only — the in-process suite does not carry one, because it does not have
the defect. Nothing is blocked either way.