Apply a function slot's domain rule once its kind is resolved, not at each merge

[dpmills]

CompactFun::merge chose how to combine two domains from the slot's KindMerge — union the alternatives for Data, take the contravariant meet otherwise — but a slot's kind is not settled until the last bound has merged, so the choice came from an answer that did not exist yet and a later bound could not undo it. Association therefore decided the outcome: an undetermined pair took the meet, {a} ⊓ {b} is {a, b}, and that deduplicated against a Data bound's own {a, b}, so a join of three collections over {a, b}, {a} and {b} typed as one collection over {a, b} in one association and was rejected in the other. The accepting association is the wrong answer: the kinds join to Data, all three domains are the data, and union_domains states the rule the meet broke — narrowing a data domain drops rows.

A positive join now accumulates the alternatives whatever the kind, and coalesce_compact_go applies the resolved kind's rule once: it meets the alternatives for a Compute reading and for a kind variable nothing pinned, and requires a single survivor for Data. The kind itself joins in the flat semilattice Unknown < {Data, Compute} < Conflict, which is the same operation at both polarities, so merge's four kind branches collapse to one. A negative merge still meets its domains directly — that rule never read the kind.

undetermined_kinds_join_without_deciding_the_domain_rule pins the exhibit in every association, and two_undetermined_kinds_join_to_a_capability_over_the_met_domain pins the join that must keep working: two undetermined-kind bounds over {a} and {b} are a capability over {a, b}, which a rule that merely rejected an undetermined join would have refused. Instrumenting the arm showed no compiled program in the suite reaches a positive undetermined join, and no inferred type changes: the suite passes forward, under DEEP_TYPECHECK=1, and under CAMBRA_REFINEMENT_ORDER=reverse.

[dpmills]

This change is part of the following stack:

• Scope inference variables with telescopes, and identify binder references by position #108
  • Hold a refined type's refinements in one unordered set instead of nested layers #77
    • Derive a Cast's refinements from its term instead of overwriting the target #78
      • Give CompactType's refinement slot the None sentinel its shape slots carry #115
        • Apply a function slot's domain rule once its kind is resolved, not at each merge #116 ◀
          • Give a function slot's domain alternatives an order-insensitive set #117
            • Reject a product with no fields instead of answering Unit #118
              • Model the type system in Lean, and check the solver's subtyping and bound merge against it #109

_{Change managed by git-spice.}