[jm-rivera]

A share whose denominator names its members gets no whole on its card. "What share of the United Kingdom's and the United States' ODA to Kenya came from the United States?" divides the US figure by the two donors together, yet the card title names only the period, so the reader cannot tell the share from the US share of all aid to Kenya. ai#79 names a whole only for named sets (DAC countries, G7) and for open axes; a column holding entities the question lists by name had no population and no entry in _populations.

This change gives such a column a population. _named_entities in services/harness/src/harness/plan_display.py resolves each named entity through the denominator's source (resolve_name) and identifies the column by its sorted graph refs, worded by entities_phrase. A name that a set of the source answers (resolve_group) stays the set's population, as before. If any name resolves to nothing, the dataset gives no named population, since the column is then unknown.

_divided_whole names a named column off the share axis only where the numerator is known to hold other entities there by identity. The numerator's own sources say what it holds. When it narrows one entity out of its frame (RequirementSide.entity), _narrowed_entity puts that entity on the column its requirement names it under (side_entities), and the frame's other columns stay as its sources hold them. A column both sides hold the same way is left out, so Kenya drops out, and the card reads "from United Kingdom and United States" in both plan forms: a separate US retrieval, or the US narrowed out of the two-donor frame. Both sides narrow alike (_side_populations), so a denominator taking the United Kingdom's row of a two-donor frame reads "from United Kingdom"; ai#79 gives such a share no whole. A side narrowed to an entity its sources do not resolve to one ref has no population: the denominator then names no whole, and the numerator shows nothing to compare. A list of names too long for the 120-character title goes to the card's note, as ai#79 does for any whole. On a pooled axis the existing rule holds, and ai#79's narrowing (_reads_whole) and ratio (_keeps_shares) rules apply to named columns as well.

Names keep the requirement's words with no article ("United Kingdom", not "the United Kingdom"). Neither the graph's labels nor planner.phrasing holds an article rule for place names, and adding one belongs with the phrasing helper, not with this change.

Frames kept from an earlier run of the plan call need no whole. _kept_asks (graph/execute.py) runs only on a stopped outcome and keeps presentable_steps, which drop every derived ops_ frame (graph/answer.py), so a kept entry holds retrieval frames only, and share_wholes gives a retrieval step no whole.

This builds on ai#81 and should merge after it.

Testing: test_a_share_of_named_donors_names_them_where_the_numerator_holds_others in tests/test_plan_display.py has seven rows over a new NAMED_DONORS source in tests/share_plan.py. named_donors and narrowed_out_of_the_frame cover both plan forms. narrowed_out_of_another_frame narrows the US out of UK and US ODA to Kenya and divides by the two donors' ODA to Kenya and Uganda, expecting the recipients named too; named_on_both_sides narrows out of a frame holding the United States as donor and recipient. Both fail when the numerator is assumed to hold the denominator's frame. same_donor and same_ref name the numerator's donor in the denominator, the second as "USA", and expect only the recipients named. value_filter filters donors to at least $1bn and expects no whole. test_a_denominator_narrowed_to_one_entity_names_that_entity expects "to every recipient and from United Kingdom" for the United Kingdom's row of a two-donor frame. ruff check, ruff format --check and ty check src are clean, and the full harness suite passes except test_server.py's zero-tools check, which fails locally only because a proxy listens on port 8083.