[3mei]

What changed and why

Every built-in probe keeps its assets beside its code, and a probe's assets resolve from the deliberate overrides, then the root its class declares. The package ships no shared assets folder.

Builds on #129.

• tool_calling in one folder (first commit). tool_calling keeps its probe, prompts, verifier, tool catalog and tests in its own folder, as the other built-in probes do: probe.py beside prompts.py and verifiers.py, assets/tool_calling/ with the catalog, its prompts and its README, and tests/probes/tool_calling/. The class declares that assets/ folder as its root. The sample_tools fixture is in tests/conftest.py, where the engine tests and the probe's tests both reach it.
• No shared assets root (second commit). The asset search path holds the deliberate overrides only (the runtime root, USERSIM_ASSET_PATH and USERSIM_ASSETS_DIR), and after them each probe reads the root its class declares. --assets-dir pins every probe to the folder it names, the same for every folder. A run that pins none records no asset folder in its simulator config, its manifest and its stored rows, and simulate --inputs reads such rows from each probe's own assets. gen-assets, validate-assets and evaluate-assets default to the domain's own bank through shipped_assets_dir(). architecture/assets.md, architecture/plugins.md, architecture/probes.md and the engine docs describe the search this way, and 01_simulate.ipynb reads each probe's own assets.

Breaking change

• The package has no usersim/engine/assets/ folder, and usersim.engine.core._assets has no packaged_assets_dir() or default_assets_dir().
• A run without --assets-dir records assets_dir: null in its simulator config, its manifest and the rows --materialize-inputs stores.
• Rows that an earlier release stored with --materialize-inputs name that release's usersim/engine/assets/ folder, so simulate --inputs refuses them and says so. Store them again with this release, or run them with --assets-dir naming a folder that holds a <probe>/ folder for each of their probes.

How it was verified

At each commit, on that commit's tree:

$ make check-all && make test && make smoke
All checks passed!
372 files already formatted
4676 passed in 92.46s (0:01:32)
All checks passed.

That is the first commit's tree. The second passes with 4675: it drops the tests of the shared folder and adds tests of the search path without it.

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

• An unpinned dry run reports where its assets come from, and a run pinned to an empty folder is refused before any model call (paths shortened):

  $ usersim simulate --locale en_US --num-rows 2 --probe-mix tool_calling=0.5,sov_ai_facts=0.5 --out out --dry-run
    probe mix:     {'tool_calling': 0.5, 'sov_ai_facts': 0.5}
    asset path:    each probe's own assets
    toolset_seed_path: .../src/usersim/engine/probes/tool_calling/assets/tool_calling/toolsets_seed.parquet
  $ usersim simulate --locale en_US --num-rows 2 --probe-mix tool_calling=1 --assets-dir empty --out out --dry-run
    - tool_calling: no asset directory at .../empty/tool_calling
  

• make health-check ran against a custom provider config for tool_calling in en_GB and for the default probe, sov_ai_dynamic in ja_JP. Both 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 tool_calling in en_GB, so the tool_use scorer read its judge prompt from the probe's folder in a process that does not import the probe:

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