Nested request payloads are not camelized

[lineoffligbot]

Generator::API#extract_body_params walks schema["properties"] exactly one
level deep, so key conversion stops at the top of the payload.

Top level converts correctly:

def query_campaign(ad_product_filter:, portfolio_id_filter: nil, ...)
  request(:post, "/adsApi/v1/query/campaigns",
    json: { "adProductFilter" => ad_product_filter, "portfolioIdFilter" => portfolio_id_filter }.compact)
end

One level down, nothing happens:

def create_campaign(campaigns: nil)
  request(:post, "/adsApi/v1/create/campaigns", json: { "campaigns" => campaigns }.compact)
end

So foo_bar becomes fooBar at the top and the caller hand-writes camelCase for
everything below. The asymmetry lands precisely on the biggest payloads. A single
campaign object has 24 properties, nested arbitrarily deep:

adProduct, autoCreationSettings, brandId, budgets, costType, countries,
endDateTime, fees, flights, frequencies, marketplaceConfigurations,
marketplaceScope, marketplaces, name, optimizations, portfolioId,
purchaseOrderNumber, salesChannel, siteRestrictions, skanAppId,
startDateTime, state, tags, targetedPGDealId

create_async_report(configuration:) has the same problem: adProduct, columns,
filters, format, groupBy, reportTypeId, timeUnit.

Options

1. Recursive key transform on the way out. Cheap, no new dependency. But it is
   lossy: targetedPGDealId round-trips to targetedPgDealId. Needs an acronym
   table (PG, SKAN, ASIN, SKU, DSP, ...). Peddler solves exactly this with
   lib/peddler/acronyms.rb feeding a Zeitwerk inflector.
2. Generate the nested types. Data classes for request bodies, the way Peddler
   generates lib/peddler/reports/*.rb. Much larger lift, but it is the only option
   where the RBS story is actually complete instead of untyped at every interesting
   boundary.
3. Do nothing, document it. Defensible for a deliberately thin generated client:
   nested bodies are raw JSON, matching the Amazon docs one-to-one.

Option 3 is the honest status quo and costs a README paragraph. Option 2 is the
real fix. Option 1 is the trap: it looks like the cheap win but silently corrupts
acronym-bearing keys unless the table is exhaustive.