feat(julia): apply_domain_weights gains category-prefix rollup on exact-key miss

When a domain_hint like "arithmetic.natural_numbers" has no exact match in
PROVER_DOMAIN_WEIGHTS, fall back to the mean of every key sharing the
"arithmetic." prefix. This lets sparse domain training data still influence
ranking through its category siblings, matching the rollup story that
VeriSim's mv_prover_success_by_class is shaped to consume.

The "no evidence → return scores unchanged" terminal behaviour and the
[0.5, 1.0] modulation range are preserved.

https://claude.ai/code/session_01YPqu7gti4azBach6ZvpRFJ

Author: claude

src/julia/api/gnn_endpoint.jl

@@ -159,9 +159,27 @@ function apply_domain_weights(scores::Vector{Float32}, domain_hints::Vector{Stri
     weights = PROVER_DOMAIN_WEIGHTS[]
     domain_rates = Float64[]
     for domain in domain_hints
+        # First: collect exact-match rates across all provers.
+        exact_matches = Float64[]
         for (_, prover_weights) in weights
             if haskey(prover_weights, domain)
-                push!(domain_rates, prover_weights[domain])
+                push!(exact_matches, prover_weights[domain])
+            end
+        end
+        if !isempty(exact_matches)
+            append!(domain_rates, exact_matches)
+            continue
+        end
+        # Miss: fall back to mean of all keys sharing the same category prefix.
+        parts = split(domain, '.'; limit=2)
+        if length(parts) == 2
+            prefix = parts[1] * "."
+            for (_, prover_weights) in weights
+                for (k, v) in prover_weights
+                    if startswith(k, prefix)
+                        push!(domain_rates, v)
+                    end
+                end
             end
         end
     end