Add GitHub-issue-driven test framework, assertion handlers, and csv-to-babeltests CLI

[gaurav]

Introduces a GitHub-issue-driven test framework for babel-validation, alongside a library carve-out and supporting tooling.

Major pieces

• New library subpackages — builds on the src/babel_validation/ carve-out that landed on main in Reorganize Babel Validation for more extensibility #101, adding sources/github/, assertions/, and tools/ alongside the existing core, services, and sources/google_sheets.
• Assertion framework — 8 AssertionHandler types (Resolves, DoesNotResolve, ResolvesWith, DoesNotResolveWith, HasLabel, ResolvesWithType, SearchByName, Needed) with an auto-generated assertions/README.md and a drift test that fails CI if it goes stale.
• GitHub integration (PyGitHub) — parses {{BabelTest|...}} wiki lines and ```yaml babel_tests: blocks from issue bodies. Open issues are expected to fail (strict xfail, so XPASS flags a closeable issue); closed issues are expected to pass.
• CSV→YAML CLI (csv-to-babeltests) for bulk-authoring assertion blocks, reusing the same handlers and YAML schema as the issue parser.
• CI workflow running pytest -m unit on PRs.
• xdist parallelism with FileLock-coordinated CSV/issue caches; the controller refreshes caches, workers share them.

Review fixes — round 1

Addressing the code review on the original revision:

• pytest -m unit now runs fully offline. Both the GitHub issue tests and the Google Sheet / blocklist tests previously fetched from the network at collection time, because their parametrization is built before pytest applies -m deselection. The marker expression is now evaluated up front (shared tests/_pytest_helpers.deselected_by_markexpr), and the GitHub fetch and Google Sheet downloads are skipped when the test would be deselected. Net effect: pytest -m unit drops from ~2.5s + network to ~0.3s offline. Row IDs (row=N) and xfail marks are unchanged, so -k 'row=42' still works.
• ResolvesWith / DoesNotResolveWith now enforce their documented 2+ CURIE arity. A single-CURIE param_set previously passed (or failed) vacuously; it now fails loudly with a clear message, matching HasLabel/SearchByName. Adds unit coverage.
• SearchByName top-N is read from targets.ini (NameResXFailIfInTop) instead of being hardcoded to 5, so it's configurable per environment.
• CachedNodeNorm.normalize_curie uses .get() to avoid a KeyError if NodeNorm ever omits a requested CURIE, consistent with the rest of the code.

Review fixes — round 2

A second review pass found eight issues, all fixed here (pytest -m unit: 49 → 56 passing).

Failures that were being hidden

• The open-issue xfail no longer swallows infrastructure errors. xfail(strict=True) with no raises= treats any exception as the expected failure. Since the marker is applied before the assertion loop and most tracked issues are open, a NodeNorm outage, an HTTP 500, or a bug in our own parsing reported XFAIL — a fully green run across every open issue. Now raises=AssertionError; the imperative pytest.xfail() calls are unaffected.
• DoesNotResolveWith no longer passes vacuously. normalize_curies returned NodeNorm's response as-is, so a CURIE omitted from it was simply absent from the mapping the handler iterates. The result dict is now seeded with every requested CURIE, so "one entry per input" is an invariant rather than an accident of the pre-warm call order.

Crashes on realistic inputs

• YAML param values are coerced to str. _to_list validated the container type but not element types, while all eight handlers expect list[str]. HasLabel: - [CHEBI:16480, NO] parsed to [..., False] (YAML 1.1 reads bare NO as boolean) → AttributeError; a numeric identifier → TypeError from _CURIE_RE.match. Numbers now stringify; bare NO/YES/ON/OFF and nulls are rejected with a quote-it hint, since the original spelling is unrecoverable by then.
• SearchByName handles a node with no label (.get('label', '') — HasLabel already guarded this).
• ResolvesWithType handles a node with no type, using the first_type() guard added in add0848.

Smaller

• ResolvesWith failure messages are deterministic. The "but expected X" id came from dict order, which depends on set iteration and cache state, so the same failing issue reported different expectations across runs. It now follows param order.
• csv-to-babeltests no longer POSTs unvalidated values. The batch pre-warm skipped the _CURIE_RE filter that test_with_nodenorm has applied since e322040, so a blank --equivalent-curie-column cell was sent to NodeNorm.
• A stale cache owned by another user no longer aborts the session. the unlink helper (unlink_if_exists after the Reorganize Babel Validation for more extensibility #101 merge) caught only FileNotFoundError; on a shared box or self-hosted runner, PermissionError escaped pytest_configure and no test ran.

New offline coverage: tests/test_assertion_robustness.py (sparse nodes, omitted CURIEs, message determinism) and three YAML-scalar cases in tests/github_issues/test_system.py.

Unrelated changes riding along in this PR

These are intentional and correct, but not part of the GitHub-issue test framework itself:

• pytest bumped 8.4.2 → 9.0.2 (plus new deps: pygithub, pytest-xdist[psutil], pytest-subtests, click, pyyaml, tqdm, filelock, python-dotenv).
• tests/targets.ini [ci]/[ci-es] restructure: both CI targets now point NodeNorm at nodenorm-es.ci.transltr.io (the old biothings.ci.transltr.io/nodenorm/ URL is gone); [ci] pairs the ES NodeNorm with the non-ES NameRes, [ci-es] with the ES NameRes. A Repositories list (scanned for embedded assertions) is added to [DEFAULT].

---

Merged main (#101)

main has since landed its own library reorg in #101, which touched the same shared
files. Merged in and resolved by taking main's structure and re-applying this branch's
work on top:

• services/nameres.py, services/nodenorm.py — main's docstrings, Protocol
  interfaces, and the clear_curie/delete_query → invalidate_curie/invalidate_query
  renames. The normalize_curies result seeding from the round-2 fixes is re-applied,
  and the "one entry per requested CURIE" guarantee is now documented in the docstring.
• sources/google_sheets/google_sheet_test_cases.py — main's cache_ttl_seconds
  constructor parameter and relative import. Nothing referenced the removed
  CACHE_TTL_SECONDS class attribute.
• tests/conftest.py — main's unlink_if_exists rename, keeping the OSError widening,
  the issue-cache cleanup, and the selected_github_issues fixture.
• CLAUDE.md, pyproject.toml — this branch is a superset of main's version.
• uv.lock — regenerated with uv lock.

pytest -m unit: 56 passed post-merge. Note the two renames for anything downstream:
clear_curie → invalidate_curie, delete_query → invalidate_query.

Follow-up: library carve-out (stacked PRs)

This PR is being kept as the base for two smaller, stacked follow-up PRs that make the GitHub-issue test framework usable as an installable library (so the Babel pipeline and Babel Explorer can consume it). They branch off this one rather than main, since the library code only exists here:

• Library packaging: import as babel_validation + split runtime/dev deps #95 — Library packaging (base: add-github-issue-tests): rename the import package src.babel_validation → babel_validation, ship src/babel_validation as the top-level wheel package, and split runtime vs dev dependencies (PyGitHub stays core). No behavior change.
• Library API: Assertion object model + run_issue_tests reporting API #96 — Library API (base: library-packaging, stacked on Library packaging: import as babel_validation + split runtime/dev deps #95): turn GitHubIssueTest into a plain Assertion object, and add a pytest-independent run_issue_tests() reporting API (IssueReport/ResultRecord) with the pytest suite rewired as a thin adapter.

Recommended review/merge order: this PR (#67) → #95 → #96. Out-of-scope library-usability improvements surfaced along the way are tracked in #97, #98, and #99.

[gaurav]

Recommended merge order: #67 → #95 → #96.

These are stacked PRs, each based on the previous branch (not main):

• Add GitHub-issue-driven test framework, assertion handlers, and csv-to-babeltests CLI #67 (add-github-issue-tests → main) — this PR.
• Library packaging: import as babel_validation + split runtime/dev deps #95 (library-packaging → add-github-issue-tests) — library packaging / rename.
• Library API: Assertion object model + run_issue_tests reporting API #96 (library-runner → library-packaging) — Assertion object model + run_issue_tests() API.

Merging in this order keeps each diff small and reviewable. Once #67 merges, GitHub will retarget #95 onto main automatically (and #96 onto library-packaging); review/merge #95 next, then #96. Out-of-scope follow-ups are tracked in #97, #98, #99, and #100.

[gaurav]

To make this reviewable, I've split it into a stack of smaller PRs (review/merge top-down):

1. Reorganize Babel Validation for more extensibility #101 — Library reorg: carve tests/common into src/babel_validation (base main)
2. Assertions framework for BabelTest expectations #102 — Assertions framework (base Reorganize Babel Validation for more extensibility #101)
3. Run BabelTest assertions embedded in GitHub issues as pytest tests #103 — GitHub-issue parsing + harness (base Assertions framework for BabelTest expectations #102)
4. csv-to-babeltests CLI for generating BabelTests YAML (4/4) #104 — csv-to-babeltests CLI (base Assertions framework for BabelTest expectations #102, independent of Run BabelTest assertions embedded in GitHub issues as pytest tests #103)

Each was reconstructed from this branch's final tree (path-based for 1–3; #104 cherry-picks the CSV commits with authorship intact) and verified offline. Plan is to merge these into main, then rebase this branch — it should reduce to little/nothing. Keeping this PR open until then.