Fourteen items across two rounds. Nothing here changes how a grounded
answer is produced or graded — the extractive path, the threshold, the
citation guarantee, and the audited evidence bundle are exactly what
0.1.0 shipped. What changed is everything around it: the tools an
operator uses to run this against a real corpus, the deployment surface
(auth, rate limiting, embedding, analytics, a real handoff to a person),
one new compliance-readiness document, and — the second round — the
infrastructure and evidence needed to trust a release: a screen-reader
test script, a Windows/macOS CI matrix that found and fixed a real
cross-platform bug in how this project's own byte-for-byte checksums
are computed, a PyPI + container release workflow proven continuously
in CI before it has ever published anything for real, and a pilot
against six real, unedited government pages that answers the one open
research question this project was carrying — whether to add semantic
retrieval — with a measurement rather than a guess.

Verified at this tag: 602 tests, ruff clean, mypy clean, 93% branch
coverage, gate PASS (14/14 suites).