Sidecar: align /predict response with Rust guidance contract (add score, avoid silent defaults)

[sheawinkler]

Problem

• Rust SidecarGuidance parser expects keys like score/confidence/expected_return/(optional multipliers).
• The sidecar currently returns prediction (class 0-4), confidence, expected_return, and probabilities, but not score. Missing keys can cause Rust to fall back to defaults (e.g. score=0.5) without operators noticing.

Goal

• Prevent ‘defaults forever’ behavior by making the contract explicit and stable.

Proposed changes

• Extend /predict response to include a guidance-friendly scalar score in [0,1] derived from ensemble probabilities (e.g. score = P(buy)+P(strong_buy)).
• Keep existing fields (prediction class, probabilities) for transparency.
• Include metadata fields that make it obvious what mapping was used (e.g. metadata.score_definition, metadata.class_prediction).
• Decide whether to also emit size_multiplier/risk_multiplier/kelly_fraction from sidecar (likely keep sizing in Rust; emit only score + confidence + expected_return unless compatibility demands more).

Acceptance criteria

• /predict includes score (0..1) + confidence + expected_return explicitly, and documentation explains mapping.
• Sidecar smoke test extended to validate the presence of score and schema metadata.

Files

• project/src/api/fastapi_server.py
• docs/SIDECAR.md (or similar)