sift-light 1.1.1-7

Changes

• Answer support now includes bounded source context around the retained exact range, including same-line qualifiers. supportContextLines accepts 0–20 and defaults to 2.
• Exact locations, both independent retrieval lanes, inspection and source validation remain available.
• Evidence scope is explicit and reflects the retained result and selected retrieval lane.
• Reuse current source checks and reads within each support request; reject oversized evidence without silently truncating it and preserve cancellation.
• Add public regression tests to CI and npm release checks.

Validation

• 634 repository tests passed, 4 skipped, 0 failed; 14 public regressions passed.
• Format, lint, type checks, generated-artifact checks, Node smoke and npm publish dry run passed.
• Pi and OMP each completed eight real-host tool cases including local model inference. Codex and Kimi MCP pagination and model-visible evidence tests passed.
• Six controlled cases were also sent to the actual Jev model with the same question, claim, one-item selection and 4,000-character upper budget. Adding context corrected one false support decision, resolved three insufficient-evidence cases as contradicted, and preserved two supported controls. Actual input token use was higher with context; this small sample is not a general accuracy guarantee.

Answer support remains opt-in and requires an enabled local-compatible HTTP JSON or Jev provider. Ordinary exact search stays model-free.