Serve the OpenAPI document at /openapi.json, as NodeNorm Redis does - #49

…release-name parser (#76)

This PR adds regression tests for six NodeNorm ES bugs. Four were fixed
before this PR, so their tests guard against a return. The other two
were found on 2026-09-29 while tracing why ORION's build against
NodeNorm ES normalized differently from its build against Redis
(biothings/NodeNormalizationAPI#40, #41); both are closed upstream but
not yet deployed to `ci`. It also adds `/status` checks that catch a
deployment reporting no backend, a malformed Babel release name, or a
broken release-notes link.

The checks these tests share now live in the `babel_validation` package
rather than under `tests/`, so Babel release names can be parsed and
compared by anything that installs it.

Best reviewed alongside #153 (NodeNorm ES response shapes against its
OpenAPI document; the two collide only in `CLAUDE.md`'s Testing
Patterns) and #157 (NameRes `/synonyms`, and `ci-nameres-es` on the
dashboard; collides only on an import line). #156, stacked on this one,
adds a tool for diffing two NodeNorm targets, to be reviewed with #148.

## Calls worth overruling

- **Open bugs are strict xfails keyed on the target's backend.**
`test_issue_40` and `test_issue_41` pass on Redis and xfail where
`targets.ini` says `NodeNormBackend = elasticsearch`; when the fixes
reach `ci`, the unexpected passes say the markers can come out. The
other four bugs are fixed, so their tests carry no xfail.
`test_nodenorm_229` uses the same setting to skip ES, which has no
`/query`, rather than probing for it.
- **Every target states its NodeNorm backend; `redis` is only the
default for a new one.** A new deployment of either kind is unlikely, a
Redis one slightly less so. Each target also states its OpenAPI path,
and `tests/unit/test_targets_ini.py` checks each backend's OpenAPI path.
That path setting can go once biothings/NodeNormalizationAPI#49 (open),
which serves the document at `/openapi.json`, reaches every ES target;
`targets.ini` says how.
- **Release-specific tests skip older deployments and fail unreadable
ones.** `require_babel_release()` skips a target serving an earlier
Babel (today `test` and `prod`, on 2025sep1). One whose `/status` cannot
be fetched, or reports no well-formed `babel_version`, fails: a skip
would make a run green that tested nothing. A successful `/status` is
cached per target; a failure is not.
- **A missing `/status` field fails rather than skips**, since every
current NodeNorm release reports `backend`, `babel_version` and
`babel_version_url`. `test_status_backend` still skips a target with no
`backend`, naming `test_status_backend_is_known` as the one that reports
it.
- **`babel_version_url` is never fetched as given.** The test builds the
URLs it would accept from the validated `babel_version` (`main` or
`master`, `releases/<v>.md` or the newer `releases/<v>/README.md`,
NCATSTranslator/Babel#1021) and fetches its own copy. Dropping the old
shape is #147.
- **The expected infores depends on the backend:** `infores:nodenorm`
from ES, `infores:sri-node-normalizer` from Redis
(biolink/information-resource-registry#140).
- **The shared code is split by what it needs.**
`src/babel_validation/core/babel_release.py` recognises and orders
release names with the standard library only.
`src/babel_validation/services/service_checks.py` holds the checks that
fail or skip a test; they check a response's shape before reading it and
`repr()`-escape whatever a service returns.

## What's here

Regression tests in `tests/nodenorm/by_issue/biothings/`, run against
every target: biothings/NodeNormalizationAPI#10 (an HTTP 500 under
`drug_chemical_conflate`), #11 (duplicate `equivalent_identifiers`,
checked for duplicates only, since its counts were one release's
cliques), #14 (a 500 from one pair of CURIEs, queried together and
apart), biothings/pending.api#338 (`NCBITaxon:2` null beside
`MESH:C029371`), #40 (`Food.txt` and `CellLine.txt` never loaded) and
#41 (a duplicate clique leader's two types merged into one list).

`tests/nodenorm/test_nodenorm_api.py` adds the `/status` checks;
NodeNorm ES has reported `1.9` and `VERSION.txt` as its `babel_version`,
both unit-tested.
[`tests/nodenorm/by_issue/biothings/CLAUDE.md`](https://github.com/TranslatorSRI/babel-validation/blob/c995d1b/tests/nodenorm/by_issue/biothings/CLAUDE.md)
records how to investigate ES against Redis, and to recommend closing an
issue whose test passes while it is still open. `CLAUDE.md` documents
the new modules, what belongs in `tests/nodenorm/by_issue/`, and `data/`
as scratch space.

## Status at merge

On `fee6ebb`, the changed tests against prod, test, ci, ci-nameres-es,
dev and exp: 185 passed, 36 xfailed (#40 and #41 on the two `ci`
targets), 42 skipped (release-gated tests on prod and test, #229 on ES),
7 failed. All 240-odd unit tests pass. None of the failures is this
PR's:

- **NodeNorm `test_openapi_json`, dev and exp (Redis):** the `max_edges`
schema bug, NCATSTranslator/NodeNormalization#409.
- **NodeNorm `test_openapi_json`, prod (ES 1.0.0):** an invalid document
still declaring `infores:sri-node-normalizer`; ES 1.2.0 on `ci` and
`test` passes.
- **NameRes `test_openapi_json`, ci, test and dev (v1.7.0):** no
`x-translator`; fixed by NCATSTranslator/NameResolution#300, in the
v1.7.1 pre-release.
- **NameRes `test_reverse_lookup`, ci-nameres-es:** the ES NameRes has
no `/reverse_lookup`; #157 moves the test to `/synonyms`.

Nothing blocks merging. Deferred: #147. The Babel-side causes of #40 and
#41 are NCATSTranslator/Babel#276 and NCATSTranslator/Babel#1123.

🤖 Generated with [Claude Code](https://claude.com/claude-code)