Constrain vision output with a JSON Schema so per-field confidences survive

[gavinbee]

Summary

Fixes the all-0.5 confidences in vision-path output (#45).

The vision call used format="json", which only guarantees valid JSON — not our nested {value, confidence} shape. Smaller models (qwen2.5vl:3b) collapse each field to a bare scalar, so _parse_field defaulted every confidence to _DEFAULT_CONFIDENCE (0.5), silently destroying the signal that drives low-confidence review.

• Build a JSON Schema from the canonical src/schema constants (_build_json_schema) and pass it to Ollama's format= so decoding is constrained to the object shape. Schema is generated from the same field lists as the prompt skeleton, so it can't drift.
• Detect an all-flat-scalar page (_response_is_flat) and log a warning instead of masking it — so this failure mode is visible if a model still ignores the schema.
• Document the model floor in docs/models.md: confidence quality is model-dependent — 3B is unreliable (constant value), 7B is the practical floor for confidence-driven features.

Investigation notes

• 7B already respected the object shape under plain format="json"; with the schema its output is unchanged ({1.0, 0.9, 0.6} spread, 0.6 landing on ambiguous successful cells) — no regression.
• 3B under plain format="json" returned flat scalars → all 0.5. With the schema it emits the object shape (though its confidences flatline — a model limitation the schema can't fix, hence the doc note).

Test plan

• pytest -q — 286 passed
• 7B end-to-end (--no-cache) — no content/confidence regression
• New tests: TestJsonSchema (schema built from canonical fields), TestFlatResponseWarning (flat detection + warning), updated format assertion

Closes #45

🤖 Generated with Claude Code

[gavinbee]

LGTM