Fix garbage DDx hypotheses when ARLL model outputs prose instead of JSON

[Copilot]

When processing non-chest images (e.g. a hand radiograph), the ARLL model outputs unstructured prose instead of JSON. The regex fallback in _parse_arll_output then captures sentence fragments as diagnosis names — e.g. "with probability", "t the spatial attention map has a attn of" — which pass _is_clinical_hypothesis and pollute the DDx ensemble with low-variance garbage, causing a false high-Sc gate pass and an empty clinical report.

Root cause

_is_clinical_hypothesis had two gaps:

1. No check that diagnosis names start with an uppercase letter (all real medical conditions do; lowercase-initial strings are always regex-captured sentence fragments)
2. _NON_CLINICAL_SUBSTRINGS missing patterns produced by reasoning-prose leakage: "attention map", "arll", " probability", etc.

Changes

rmoe/agents.py

• _is_clinical_hypothesis: Reject any candidate whose first character is not uppercase. This single rule eliminates all observed garbage fragments.
• _NON_CLINICAL_SUBSTRINGS: Add "attention map", " attn ", "arll", " probability", "how to", "the model", "approach this" as defence-in-depth for uppercase-starting false positives.

# Before: "with probability" and "t the spatial attention map…" passed the filter
def _is_clinical_hypothesis(name: str) -> bool:
    if len(name.strip()) < _MIN_DIAGNOSIS_LENGTH:
        return False
    low = name.lower()
    return not any(s in low for s in _NON_CLINICAL_SUBSTRINGS)

# After: lowercase-initial strings are immediately rejected
def _is_clinical_hypothesis(name: str) -> bool:
    stripped = name.strip()
    if len(stripped) < _MIN_DIAGNOSIS_LENGTH:
        return False
    if not stripped[0].isupper():   # sentence fragments start lowercase
        return False
    low = stripped.lower()
    return not any(s in low for s in _NON_CLINICAL_SUBSTRINGS)

When all candidates are filtered, _parse_arll_output returns an empty ensemble, causing ReasoningExpert.execute() to fall through to _fallback_ensemble — a structured output rather than garbage.

engine.py

• Remove unused imports (CYAN, MAGENTA, WHITE, _kv) and fix a spurious f-string that were causing CI pyflakes failures on main.

tests/test_agents.py (new)

• 21 tests covering _is_clinical_hypothesis (valid diagnoses, all five garbage fragments from the reported run, edge cases) and _parse_arll_output (clean JSON, pure-prose input, mixed input, JSON with garbage diagnosis fields).

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.