createUser() drew a random mobile and recovered from a collision by catching
the unique-constraint violation and calling resetManager(). That hands back a
brand-new EntityManager, which detaches every entity the running test had
built so far; its next flush died with "Multiple non-persisted new entities
were found", always in a different test and never reproducible in isolation.
The number is now checked before the insert, so the collision never reaches
the database and the manager stays open.
testParentIdAddsNoQueryPerSpecialty counted queries on the first request of
each size, so one-shot per-process caches — site config, subscription plan,
Doctrine metadata — landed inside the count or not depending on which tests
had run before it. Both requests are now warmed first; the assertion measures
steady-state growth, which is what it was always about.
The 23 PHPUnit notices were all one complaint: doubles created with
createMock() that never had an expectation. The ones that only stub return
values became createStub(); in SmsServiceLookupOnlyTest the provider and the
bus got the expectations they were missing, since "dispatch does not touch
the provider" and "sendNow does not enqueue" are exactly what that suite is
there to prove.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`GET /api/v1/doctors` could not answer either question the public search box
asks. Typing a specialty name returned nothing, because `name` only matched
`d.name`. And `specialty_id` matched one id exactly, so a parent group only
found doctors who happened to carry the parent — which they usually do, but
only as a side effect of `expandWithAncestors` running on save. A doctor
imported through any other path has no denormalised parent, and a search
guarantee resting on a save-time side effect is not a guarantee.
`expandWithDescendants` mirrors the existing ancestor walk over the same cached
parentMap, so no extra query. It deliberately keeps unknown ids instead of
dropping them like its mirror does: the result feeds an `IN (...)`, and an empty
array turns the filter into a no-op that returns every doctor — an unknown id
must mean "nothing", never "everything".
Both specialty filters use their own EXISTS alias rather than the shared `s`
join. Two conditions on one alias force a single join row to satisfy both, so a
doctor filtered by specialty A while searching the name of specialty B was
silently dropped. Verified by reverting to the shared alias and watching
testFilterOnOneSpecialtyWhileSearchingTheNameOfAnother fail.
toListArray now carries specialties[].parent_id so a client can tell the main
specialty from a sub-specialty instead of printing all of them. It is a string,
matching toDetailArray and the sibling `id` key — one concept should not have
two types across two endpoints. Reading the id off the parent proxy costs no
query; measured 6→11 queries with four more doctors both with and without the
field. That growth is a pre-existing N+1 (findWithFilters does not fetch-join
specialties, unlike findByClinic) and is left untouched here.
Also drops the phantom `search` parameter from the OpenAPI annotation — it was
advertised but never read, so a client sending it got an unfiltered list — and
documents the six live parameters that were missing.
Note for deploy: DoctorRepository gained a constructor argument, so a stale
container fails with ArgumentCountError until cache:clear runs.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
