Added
- Benchmark + quality harness (
benchmarks/bench.py, stdlib-only): scores the summarizer's salient-sentence selection against lead/tail/random baselines and verifies both backends select identical sentences.--checkgates these
in CI on both the numpy and pure-Python paths.
Fixed
- Deterministic summaries across backends — ranking rounds away sub-epsilon float noise and breaks ties by position, so a committed
context.mdno longer depends on whether numpy is installed. - Cleaner "Next steps / open threads" — drops echoed prompts/instructions, trivial questions, header lines, and Recall's own injected boilerplate; bare questions must be substantive.
Full changelog: https://github.com/raiyanyahya/recall/blob/master/CHANGELOG.md