feat: spill Pydantic BaseModel tool results

[thomwebb]

What

The spill plugin now also bounds Pydantic BaseModel tool results, not just dicts.

This closes the plugin's biggest practical gap: most big-output built-in tools return models, not dicts — agent_run_shell_command → ShellCommandOutput, list_files → ListFileOutput, invoke_agent → AgentInvokeOutput — and until now those all passed through spill untouched. (The dict-returning tools, e.g. create_file's {"path", "message"}, were the small ones.)

How (small on purpose)

• _field_map() — a dict(result) view for models; the existing planner/notice/store machinery is reused verbatim, operating on field names instead of dict keys.
• _commit_replacements() — item assignment for dicts, setattr for models; on a rejected assignment (frozen model, validate_assignment validator) it restores already-applied fields and re-raises, and the outer hook's existing fail-open path keeps the result byte-identical inline.
• No allowlists, no contract pinning, no new config: any BaseModel with oversized plain-string top-level fields is eligible; exotic cases (frozen, rejecting validators) simply stay inline.

Verified end-to-end against current app main: a ShellCommandOutput with 4 KB stdout at a 500-byte cap becomes head/tail preview + retrieval notice, file written verbatim.

Tests (tests/test_spill_models.py, local models — no app-tool imports)

• oversized field spilled in place; file verbatim; notice contains path; other fields untouched
• under-cap model untouched
• frozen model fails open, stays inline
• commit rollback restores prior fields when a validate_assignment validator rejects a preview

Full spill suite on app main: 21 passed, 4 skipped (the 4 skips are the per-agent tests gated on the runtime execution-context seam, from #9). Ruff clean.

[WSxDemise]

[Wes's CodePuppy Agent Review]

Adds mutable Pydantic BaseModel handling to spill, with documentation and focused unit coverage.

Qodo Merge Triage

Qodo: not configured for OSS/public GitHub.

Really Should Fix These

None.

Nits

None.

Scope observations — not changes requested on this PR

• [P0] [UNIMPLEMENTED] Hook-context BaseModel results remain unbounded — code_puppy_core_plugins/code_puppy_agent/SYSTEM_PROMPT_CONFIG_AND_I18N.md:83
  • Impact: When Emoji Filter strips an emoji from a shell command (on by default), core converts ShellCommandOutput to str before spill, leaving large stdout in model context.
  • Fix: File a companion code_puppy core PR to preserve structured results while prepending hook context, with an integration test for an emoji-filtered shell result.