Experiment resume can skip changed configurations due to run_id collisions

[rschumann]

Description

ExperimentRunner.run_single() uses run_id to decide whether a run has
already completed. The current ID contains the model alias, regime, condition,
and seed, but not the requested case count or other result-affecting
configuration.

For example, these two configurations currently produce the same ID:

from mech_gov.experiment.runner import RunConfig, _make_run_id

for n in (10, 1000):
    config = RunConfig(
        model_name="mock",
        regime_name="R2",
        condition="S0",
        seed=42,
        cases_per_condition=n,
    )
    print(
        n,
        _make_run_id(
            config.model_name,
            config.regime_name,
            config.condition,
            config.seed,
        ),
    )

Output:

10 mock-R2-S0-seed42
1000 mock-R2-S0-seed42

Because the completion check happens before execution, skip_completed=True
can treat the changed experiment as already completed.

Relevant code:

• _make_run_id() in src/mech_gov/experiment/runner.py
• the completion check near the start of ExperimentRunner.run_single()

Proposed direction

Would you be open to separating the readable run label from a deterministic,
versioned identity derived from allowlisted result-affecting inputs?

A possible identity could include the case count, resolved model ID, condition,
regime, seed, semantic distribution fingerprint, and any explicitly supported
result-affecting parameters. Existing JSONL records would remain untouched.

Before submitting a patch, I would like to confirm your preferred compatibility
behavior for legacy records:

1. Warn and execute when only a legacy label matches; or
2. Preserve legacy label-based resume behind an explicit compatibility option.

I can adapt the patch to the agreed approach and include offline tests for
collision prevention, determinism, legacy records, and secret exclusion.