chore(models): regenerate against schema 0.2.0 + close the drift

[tbsvttr]

Stacked on #6.

Closes the drift Scott flagged in the initial review. models.py here has been on schema line 0.1.0 since the port; content repos ship 0.2.0. Every end-to-end validation test was xfailed against datasworn_version mismatches and downstream shape drift. This PR proves out the "regenerate + post-process" workflow: fetch the current distribution schema, run datamodel-code-generator, run scripts/post_process_models.py, and everything lines up.

What's in here

• Regenerated models.py against https://raw.githubusercontent.com/datasworn-community/datasworn/main/packages/core/json/datasworn.schema.json (v0.2.0).
• New post-processor pass feat(models): convert RootModel[str] wrappers to TypeAlias + resolve discriminated-union bases #5: leading-underscore JSON keys (_id, _source, _comment, …) get renamed to field_id etc. by datamodel-code-generator to avoid colliding with pydantic's private namespace. Added a pass that renames them back to the natural attr name (dropping the field_ prefix) while keeping Field(alias='_X') intact — so .id reads naturally without breaking JSON validation. 126 attrs renamed.
• Handle the compact single-line RootModel[str] shape. Codegen emits a compact one-line form for description-only fields (e.g. CssColor); the multiline regex missed those. Added an inline variant.
• Updated post-processor docstring to describe all five bugs it handles (previously said "two").
• Renamed the ironsmith/starsmith xfail group to test_community_known_content_bugs with a clear reason — those failures are option [key] values with spaces ("area of expertise", "define the dark power in you", "true name") that don't match the DictKey pattern ^[a-z][a-z0-9_]*$. Content-quality bugs to file against datasworn-community/community-content, not drift.
• Rewrote test_root_model_str_wrappers_gone to inspect the source (re.findall(r\"^class \\w+\\(RootModel\\[str\\]\\):\", source)) instead of runtime shape. Runtime check tripped on ConditionMeterKey(RootModel[DictKey]) because DictKey post-processes to Annotated[str, …] and pydantic collapses it to `str`. Source-level check matches what the post-processor actually operates on.
• README refresh: replaced the "known drift ⚠️" warning with a schema-line status, swapped the regen command to fetch the distribution schema (not the source schema), and updated step 3 to reflect that a green test run now confirms the regen.
• .gitignore: added tests/.cache/.

Post-processor stats after regen

• 3 date-field pattern strips
• 82 RootModel[str] → TypeAlias rewrites (multiline + compact)
• 5 discriminated-union bases resolved
• 126 field_X → X attribute renames
• 3 Features/Dangers/Denizens stub sites rewritten

Test results

```
tests/test_load_rules_packages.py .......xx [ 75%]
tests/test_type_alias_ergonomics.py ... [100%]

======================== 10 passed, 2 xfailed in 0.46s =========================
```

• ✅ classic, delve, lodestar, starforged, sundered_isles
• ✅ ancient_wonders, fe_runners
• ❎ ironsmith, starsmith — xfail on content-quality bugs (not drift)
• ✅ all 3 ergonomics tests

Test plan

• Regenerate models.py cleanly from schema 0.2.0
• Post-processor idempotent on second run (re-run reports "no changes")
• uv run pytest tests/ green: 10 passed, 2 xfailed, no failures
• CI green on this branch
• Follow-up: file content-quality bugs against community-content for ironsmith + starsmith