1. `stat_key` alone (no `pair_key`, no `stat_key_a`/`stat_key_b`): ranks
   every area at one geography level by its latest figure for that stat,
   for one country -- e.g. "which counties have the highest median sale
   price" (country="ireland"). `stat_key` comes from `query_data`'s
   dataset-listing mode, for the SAME country. `level` omitted uses this
   ranking's own default level; pass one of that dataset's own
   `compatible_levels` for a different one -- a level this ranking
   doesn't have registered returns an empty list rather than an error.
   Where the underlying stat has no honest per-area denominator (crime,
   homelessness, live_register and similar -- StatsMapped's own
   RANKING_NO_DENOMINATOR_STATS), each row's `rate_per_1000` is the real
   figure to rank/compare by, not `latest_value`, which is a raw count
   dominated by area population size. Always carry forward every entry
   in `caveats` when using a row in an answer.
2. `pair_key` alone: full detail for one registered comparison pair --
   each axis's label, unit and publisher, the correlation stats (r, rho,
   and a leave-one-out sensitivity range naming the single most
   influential area), and caveats. `pair_key` comes from mode 4's own
   response, for the SAME country.
3. Both `stat_key_a` and `stat_key_b` given: does StatsMapped have a
   registered, hand-vetted comparison between these two stats? Registry-
   backed only -- never computes a fresh correlation for an arbitrary
   pair. Both stat_keys come from `query_data`'s dataset-listing mode,
   for the SAME country. `verdict` is one of `"SUPPORT"` (a real,
   hand-vetted registered pair with no open caveats -- may be treated as
   a confirmed relationship), `"QUALIFY"` (hand-vetted, but the evidence
   carries real caveats -- e.g. no robustness check for outliers, or an
   unverified geography-level join; read `uncertainty` and `reasons`
   before presenting it as confirmed), `"REJECT"` (a real structural
   impossibility or a human-vetted "no" -- the two stats share no
   geography level at all, or a reviewer rejected this exact pairing),
   or `"INSUFFICIENT"` (not registered, not ruled out either --
   StatsMapped genuinely hasn't vetted this pair; never treat this as
   "probably comparable"). `uncertainty` names 4 separate dimensions
   (data_quality, comparability, statistical_strength,
   causal_strength) -- `causal_strength` is always `"not_established"`,
   since no comparison here implies causation regardless of verdict.
   `comparable` (DEPRECATED, kept only for callers that haven't
   migrated) collapses `verdict` to the old 3-way yes/no/unknown --
   `"yes"` for both `SUPPORT` and `QUALIFY` (both mean "hand-vetted",
   the old `comparable` meaning this field has always carried; the
   caveats a `QUALIFY` pair carries live in `uncertainty`/`reasons`,
   not in demoting `comparable`), `"no"` for `REJECT`, `"unknown"` for
   `INSUFFICIENT`. Prefer `verdict` directly when you need to
   distinguish a fully-confirmed `SUPPORT` from a caveated `QUALIFY`.
   Read `reasons` before presenting any answer other than `"SUPPORT"`
   as unqualified.
4. None of the above given: lists every registered cross-dataset
   comparison pair for one country -- e.g. "median sale price vs new
   dwelling completions per 1,000 residents". A small, hand-curated set,
   not an arbitrary-pair engine: pass one of the returned `pair_key`
   values to mode 2 for the real correlation and axis detail.

   `level` is only meaningful together with `stat_key` (mode 1); giving
   it without `stat_key` raises rather than silently dropping it and
   falling through to mode 4's unrelated pair listing.