Release 0.5.1: EvalBench run reader, canonical rubrics, single-input configurator

[caohy1988]

Summary

Version bump 0.5.0 → 0.5.1 plus the changelog cut for everything merged since v0.5.0 (2026-08-11), 17 commits.

In the wheel: EvalBench BigQuery run reader — new bigquery_agent_analytics.evalbench module (#444); canonical evaluation rubrics in core — new bigquery_agent_analytics.evaluation_rubrics module extracted from quality_report (#430/#431); golden Q&A matching producer — new bigquery_agent_analytics.golden_matching module with embed_texts, DEFAULT_GOLDEN_THRESHOLD, and match_golden_qa exported top-level (#428/#432); CLI judge-feedback escaping and max-length fix (#438); example fixes (#426/#427).

Repo/live-template side (labeled dashboard sections in the changelog, not in the wheel): the single fully-qualified table-ID configurator entrance with its fail-closed field machine, two error classes, Enter bridge, and nine-fixture browser smoke (#448/#449, superseding #403's separate-fields decision); Console table-link paste (#424); the attested external-access contract — two-control external_access_verification, durable incident record, evidence-consistent status tests, and the weekly staleness workflow — born from the real external-user failure in #445 (#446); the end-user manual (#425); Grafana distinct-span call counting, half-open 72-hour windows, and hardened sync lint (#433) plus the one-command local run (#421/#422).

Docs/infra: LangSmith re-export + 24h ingest window (#423), thinking-token accounting (#441), ontology concept-index semantics (#439), time-series point markers (#437), and the pyink<26 CI pin (#447).

Version note

Three of the wheel changes add new public modules (evalbench, evaluation_rubrics, golden_matching), which strict SemVer could read as MINOR. Decision: this release ships as 0.5.1 (release owner's call, 2026-08-29).

Release procedure after merge

Publish a GitHub release tagged v0.5.1 on the merge commit — release.yml verifies the tag against pyproject.toml and publishes to PyPI via trusted publishing. (Producer tracing-v* releases are unaffected; the workflow's v* guard separates them.)

Verification

• No code changes — pyproject.toml version line + CHANGELOG.md only; no other 0.5.0 references exist in src/ or packaging files.
• Dashboard suite 29/29 at this head.
• CI's Build-package job re-verifies the version against the built artifacts.

[caohy1988]

P1: Cut this release as 0.6.0 rather than 0.5.1. The changelog explicitly says the project follows SemVer, and the merge result adds three public SDK surfaces: evalbench, evaluation_rubrics, and golden_matching, with rubric and golden-matching functions also re-exported from the package root. The v0.5.0 release established the same project precedent by using a MINOR bump for a new export subpackage and CLI surface. Publishing these additions in a PATCH band makes the immutable PyPI/version contract inconsistent with both policies. Please change this line and the changelog heading to 0.6.0, then update the PR title/body and planned tag to v0.6.0.

[caohy1988]

P2: Add the golden Q&A matching API from #432 to this release section. The branch was cut at 16 commits, but main now has a seventeenth post-v0.5.0 commit before this release: src/bigquery_agent_analytics/golden_matching.py plus top-level exports for DEFAULT_GOLDEN_THRESHOLD, embed_texts, and match_golden_qa. GitHub CI builds the synthetic merge, and that 0.5.1 wheel already contains golden_matching.py, so users would receive an undocumented public module. Please add it to Release highlights and Added, and update the PR summary/count from 16 to 17 commits.

[caohy1988]

Fixed in 52fcc87. Merged main into the release branch (so the branch itself now contains 235cdd4, not just the synthetic merge) and documented the module in both Release highlights and the Added section: bigquery_agent_analytics.golden_matching with embed_texts, DEFAULT_GOLDEN_THRESHOLD, and match_golden_qa, noting it as the producer side of the #378 answer-key grounding, extracted verbatim from quality_report. PR summary updated to 17 commits, and the version note now counts three new public modules — which strengthens the 0.6.0-vs-0.5.1 question already flagged there.

[caohy1988]

Full review of PR #450 at 67794c8: request changes.\n\nFindings posted inline:\n- P1: release as 0.6.0, not 0.5.1. The merge publishes three new public SDK modules/surfaces, and both the stated SemVer policy and the v0.5.0 release precedent require a MINOR bump.\n- P2: add #432 golden Q&A matching to the changelog and release summary. The branch was cut at 16 commits, but current main adds a seventeenth pre-release commit; the synthetic merge wheel already ships golden_matching.py and its top-level exports.\n\nVerification:\n- compared every commit from v0.5.0 through current main; the other 16 commits are represented accurately\n- inspected the green CI dist artifact: METADATA reports 0.5.1 and the wheel contains evalbench.py, evaluation_rubrics.py, and golden_matching.py\n- confirmed the release workflow requires the tag to equal pyproject.toml, making the selected version immutable once v0.5.1 is published\n- all current CI checks are green\n\nAfter updating from main, add the #432 notes, change pyproject.toml and the changelog heading to 0.6.0, and update the PR title/body and planned tag to v0.6.0. No other release-note or packaging findings.