Function try_validation_probe

crates/openshell-router/src/backend.rs:507–539

Run one validation probe, retrying with its fallback body only when the upstream specifically rejected `max_completion_tokens`. That retry exists for the GPT-5+ (`max_completion_tokens`) versus legacy (`max_tokens`) chat split. Firing it for any request-shape rejection would issue a second, pointless probe when the real signal is "wrong protocol for this model", and a transient `429`/`5xx` on tha

(
    client: &reqwest::Client,
    route: &ResolvedRoute,
    probe: &ValidationProbe,
    headers: &[(String, String)],
)

/// terminal failure that stops the caller from reaching a protocol that would
/// have validated.
async fn try_validation_probe(
    client: &reqwest::Client,
    route: &ResolvedRoute,
    probe: &ValidationProbe,
    headers: &[(String, String)],
) -> Result<ValidatedEndpoint, ValidationFailure> {
    let result = try_validation_request(
        client,
        route,
        probe.path,
        probe.protocol,
        headers,
        probe.body.clone(),
    )
    .await;

    if let (Err(err), Some(fallback_body)) = (&result, &probe.fallback_body)
        && err.kind == ValidationFailureKind::RequestShape
        && err.details.contains("max_completion_tokens")
    {
        return try_validation_request(
            client,
            route,
            probe.path,
            probe.protocol,
            headers,
            fallback_body.clone(),
        )
        .await;
    }

    result
}

/// Send a single validation request and classify the response.
async fn try_validation_request(