dsh-vision-bridge v0.2.1

v0.2.1 — Multi-image evidence robustness

Improvements

• Stronger multi-image attachment separation. The multi-image analyzer prompt now explicitly requires one images[] entry per distinct attachment and forbids merging attachments into a single entry, even when they appear adjacent or visually continuous.
• Bounded output-contract retry. Multi-image Vision analysis now uses a single shared budget of at most two attempts per work unit, retrying once when the model output fails strict JSON parsing or any multi Evidence validation check. Provider/transport/stream failures are never retried.
• Resilience to envelope noise. A bounded, leading-position-only tolerance now handles the observed leading U+200B zero-width space that the provider may emit before its JSON object. The tolerance is limited to leading U+200B characters on the multi-image path only; content inside the JSON value is never altered.

Preserved guarantees

• Maximum two multi Vision attempts per work unit; there is no third attempt.
• Strict Evidence validation remains authoritative: images.length must equal the number of input images, indexes must be exactly 1..N, and relation invariants are unchanged.
• Single-image behavior is unchanged (no retry, no forced temperature, existing prompt).
• Temperature 0 applies only to multi-image Vision attempts.
• Stage 3B Evidence cache semantics are unchanged (EVIDENCE_POLICY_VERSION is now 3, so cached Evidence from earlier policy versions is never reused).
• No heuristic Evidence repair: invalid output is discarded and never forwarded downstream.

Compatibility

• Backward-compatible at the public API level; no public exports added or removed.
• Install as before: add the plugin from this release tarball to a DSH profile.