v0.4.0

Added

• retriever::Retriever trait plus VectorStoreRetriever adapter and the
  score_retriever / retrieve_all drivers, so backends that are not vector
  stores (lexical / BM25 / hybrid rerankers, remote search APIs) can be scored
  with the same IR metrics. RetrievalHarness now delegates to this driver.
• Qrels::from_beir loads a downloaded BEIR / BRIGHT dataset
  (queries.jsonl + qrels/<split>.tsv, both 3- and 4-column layouts).
• synthetic module: deterministic, seeded "needle in a haystack" corpus +
  qrels generator for reproducible benchmarks and fixture-free tests.
• RetrievalHarness::with_bootstrap and MultiReport::with_bootstrap
  attach percentile-bootstrap confidence intervals to every metric;
  MultiReport::delta_markdown renders a head-to-head delta table.

v0.3.2

Tests

• Cover ingestion graph knowledge gain interaction (#7)

v0.3.1

Documentation

• Align docs/decisions.md heading + AGENTS feature list with v0.2.0+ (#2)

v0.3.0

Build

• Drop path override on rig-memvid dev-dep so release-plz can resolve crates.io fallback

CI

• Add release-plz workflow and config

Added

• Bootstrap confidence intervals on MetricReport::mean: new MetricCi
  type plus MetricReport::bootstrap_ci(iterations, level, seed) and
  builder MetricReport::with_bootstrap_ci(...). Uses a deterministic
  SplitMix64 stream (no rand dependency) so the same seed reproduces
  the same interval. MetricCi is #[serde(default, skip_serializing_if = "Option::is_none")] so existing JSON reports
  stay schema-compatible. MetricDelta carries current_ci /
  baseline_ci alongside the existing mean delta.
• Non-zero CI exit support on ReportDiff: is_clean(&gate) -> bool
  and exit_code(&gate) -> i32 (0 = pass, 1 = regression). Lets eval
  binaries gate CI with std::process::exit(diff.exit_code(&gate))
  without rebuilding the regression-walking logic at every call site.
• observe feature: emit MultiReport and ReportDiff as
  rig-tap-compatible JSON envelopes on the rig_tap tracing target
  without depending on rig-tap. Introduces EvalEnvelope + EvalKind
  (eval.retrieval_report, eval.regression_diff),
  report_envelopes, diff_envelopes, and tracing helpers
  emit_report / emit_diff. Each event carries the full JSON envelope
  plus stable scalar rig_tap.kind / rig_tap.metric /
  rig_tap.regressed / rig_tap.conversation_id attributes so existing
  OpenTelemetry collectors route eval reports the same way they route
  prompt and tool events.
• FreshnessReport and FreshnessQueryRollup for rolling
  StalenessReport / ConflictReport outputs into MultiReport.
  MultiReport::with_freshness attaches the dataset/per-query rollup, while
  with_freshness_metrics also appends score-like
  freshness.stale_free_rate@k and freshness.conflict_free_rate@k metric
  rows so freshness regressions trip the existing RegressionGate / diff
  path.
• Opt-in MinHash-style chunk near-duplicate linting via
  NearDuplicateLintConfig, NearDuplicatePair,
  ChunkStats::near_duplicate_pairs, and
  ChunkLintWarning::NearDuplicateChunks. The default remains disabled so
  existing ingestion gates keep their current warning set until hosts opt in.