#3265 step 4: report call-resolution rates per language and per repo #3331

Part of #3265. Depends on #3328. Cheap, and it's the first honest quality number for the call graph -- it needs no outside tool.

For every scan, report the share of call sites that are internal-unique / internal-scoped / ambiguous / external, per language and per repo. Record it in the master DB, and add it to the architecture brief so it's visible on every self-scan.

Also track it across language-crucible (like tree_sitter_accuracy_history.csv), so per-language changes to the calls_out rules (e.g. the #3264 batches) show up as a measured movement instead of an unmeasured one.

Caveat to state wherever it's shown: a high unique rate means the resolver made a confident choice, not that the choice was correct. Correctness is #3332's benchmark. Per CLAUDE.md, no README/badge claim until that exists.

Done when: the rates are in the DB and the brief, plus a corpus-wide baseline per language.