[3mei]

What changed and why

Built-in probes stay self-contained: no built-in imports another probe's module, each built-in package's imports of internal usersim modules are counted and can only go down, and the capability report's refs and the probe catalog follow the registered probes and scorers.

Builds on #126.

• persona_to_tags in core (first commit). persona_to_tags, with the region, age, education and occupation tables it reads, is in core/persona_tags.py. sov_ai_facts, sov_ai_dynamic, sov_ai_multilingual_parity, safety_chat_pressure and safety_agentic import it from there, so no probe imports another probe's module for its persona tags; financial_services has its own persona_to_tags. Its tests are in tests/engine/core/test_persona_tags.py, and the sov_ai_facts tests check that its banks serve each UK nation and German Land.
• Internal imports held to their counts (second commit). tests/test_probe_private_imports.py counts, for each built-in package, the names it imports from usersim modules other than usersim.extension, usersim.testing and its own package. A package over its allowance fails, so does an allowance above the count, and so does a package with no allowance. The authoring guide describes the check where it covers registering a built-in.
• Capability refs from the registries (third commit). implementation_refs_for_capability gives each probe and scorer a capability names the file its registered class or function is defined in, so a probe or scorer that moves takes its ref along. A test checks that every probe and scorer a capability names resolves to a file.
• The catalog follows the built-ins (fourth commit). A test checks that the catalog table in docs/probes.md has one row for each built-in probe and no other.

How it was verified

At each commit, on that commit's tree:

$ make check-all && make test && make smoke
All checks passed!
368 files already formatted
4657 passed in 90.88s (0:01:30)
All checks passed.

That is the first commit's tree; the second, third and fourth pass with 4660, 4661 and 4662 tests, their new tests included.

At the head:

$ make check-license-headers
🎉 All files have correct license headers!
$ make check-doc-links
✓ every relative link, and every link into this repository's main branch, resolves
$ make check-notebooks-clean
OK 4 notebook(s) carry no outputs
$ make check-dist
✓ installed package runs from a clean environment
$ make check-extensions
all extension seams work against an installed distribution

• core/persona_tags.py holds, byte for byte, the 391 lines that leave sov_ai_facts/task_derivation.py.

• The second commit's tests fail when a package imports a name its allowance does not count, and when an allowance is above its count:

  E       AssertionError: A built-in package imports more names from internal usersim modules than its allowance. Import them from usersim.extension, or keep them in the probe's own package: {'sov_ai_facts': [...]}
  E       AssertionError: lower each allowance to its package's count, or drop a removed package, (allowed, counted): {'sov_ai_facts': (23, 22)}
  

• At the third commit, each of the 46 capabilities gets the same refs, in the same order, as at the base.

• The fourth commit's test fails with a catalog row removed:

  E         At index 7 diff: 'safety_agentic' != 'identity_disclosure'
  

• make health-check ran against a custom provider config for sov_ai_multilingual_parity in hi_Deva_IN, which picks its query by persona tags. It ended with ✓ simulate, eval and report all completed, and every model's requests show failed=0.

• usersim simulate, then usersim eval --scorers auto, ran against the same config for safety_chat_pressure in ko_KR:

  judge: scored
  language_compliance: ok (error=None)
  refusal_basics: ok (error=None)
  response_shape: ok (error=None)
  safety_chat_pressure: ok (error=None)
  

Results impact

Conversations, prompts, scores and the capability report are unchanged.