Move dry_run_role off public LLMClientConfig schema #63

Context

PR #62 added a dry_run_role: str | None = None field to LLMClientConfig in clawloop/train.py:39-42 so that clawloop run --dry-run can identify which mock client to substitute for each role. This was chosen over the prior approach (mutating LLMClientConfig.model with a marker string) on reviewer feedback (PR #62 review comment 1) because mutating model could pollute downstream consumers like _build_entropic, which reads tc.model and propagates it into entropic_cfg.

The current design works and is well-tested (tests/test_cli.py::test_dry_run_role_field_survives_pydantic_copy), but it still has a minor smell flagged in self-review: a CLI testing affordance now lives on a public Pydantic model. Every model_dump_json() of a TrainConfig emits \"dry_run_role\": null for each LLM client. JSON Schema generators surface it. Future config-validation UIs see it.

Proposal

Move the role mapping into a side-channel owned entirely by clawloop/cli.py:

# in cli.py
from weakref import WeakKeyDictionary

_dry_run_roles: WeakKeyDictionary[LLMClientConfig, str] = WeakKeyDictionary()

_install_dry_run_clients populates this dict; the patched _make_llm_client reads from it via _dry_run_roles.get(cfg). This:

• Removes dry_run_role from the public LLMClientConfig schema entirely
• Survives Pydantic model_copy() the same way the field does, as long as the original object stays referenced (it does — config.llm_clients holds it)
• Keeps all dry-run vocabulary in cli.py where it belongs

Trade-off: WeakKeyDictionary requires the cfg objects to be hashable. Pydantic v2 models are hashable by default, so this should work.

Acceptance

• LLMClientConfig no longer carries dry_run_role
• model_dump_json() of LLMClientConfig does not emit any dry-run markers
• All existing dry-run smoke tests still pass (test_run_*_dry_run_*, test_dry_run_role_*_survives_pydantic_copy)
• The Pydantic-copy test is rewritten to assert the side-channel survives copies

Related

• PR fix: address PR #60 review comments on clawloop run --dry-run + taubench #62 (this comment was raised in self-review using the A Philosophy of Software Design framework: principle 3, information hiding)
• PR chore: reintroduce clawloop run as a TrainConfig wrapper #60 review comment 1 (gemini-code-assist) — the prior round