Rebuild missing SOLR content everywhere it is read, not just in term_info

[Robbie1977]

Rebuild missing SOLR content everywhere it is read, not just in term_info

#108 fixed term_info. It turns out four hydration paths had the same silent skip, and one of them was worse than "incomplete".

The connectivity queries were returning nothing

Not fewer rows — a confident empty answer. Before and after, live:

FBbt_00110929  alpha/beta core Kenyon cell         0 rows, count 0   →  236
FBbt_00048273  wing bristle mechanosensory neuron  0 rows, count 0   →  604

152 of 4,000 sampled connectivity classes are in that state. Someone asking what is downstream of a Kenyon cell was being told, with a straight face, nothing.

Four paths, one cause

Records that reach the PDB after the bulk indexer last completed have nothing in vfb_json, and precompute live query results has not completed since 22 June.

path	reads	did
_owlery_query_to_results	anat_query, anat_image_query	skipped the id — rows and count short
_fetch_connectivity_entries	up/downstream_connectivity_query	skipped the class — empty result
_bulk_fetch_per_instance_connectivity	its own @with_solr_cache entries	skipped, warning that counts were "a slight underestimate"
get_hierarchy	term_info	skipped — and lost the whole branch beneath the term

All four now collect what SOLR returned, work out what is missing, rebuild it, and carry on.

How the rebuilds work

The first two borrow the indexer's own query, same as #108 — nothing re-implements the schema. Measured on the live PDB a rebuild costs one round trip rather than per-id work (1 id 0.37s, 50 ids 0.91s, 200 ids 0.80s, 500 ids 1.49s, 3 ms/id), so ids go in batches of the indexer's own REQUEST_BATCH_SIZE. A set larger than one batch is looped, not truncated — dropping the tail would be the same quietly-short answer this PR exists to stop. Rebuilt payloads are returned in memory and used by the current request, because the write uses commitWithin 60s.

The per-instance connectivity miss is a different animal: not a gap in the bulk index but VFBquery's own result cache never having been asked for that instance. So the fix is to ask — get_neuron_neuron_connectivity is cache-wrapped, so computing one also fills the cache. Deliberately unbounded: an individual call should fix itself even if that call is slow. Misses are rare enough for that to be safe — 0 of 4,000 connectivity individuals sampled had none, and the case that surfaced it was 1 of 794.

#108 was serving but never indexing

solr_client builds its update URL from SOLRserver/SOLRcollection and a VFBquery deployment sets neither, so get_solr_update_url() returned None and send_solr_payload returned False without making a request. Confirmed against live 1.22.49: get_term_info("VFB_00107fob") served a complete document while id:VFB_00107fob stayed numFound 0 past the 60s commitWithin. Both are now seeded from VFBquery's own pysolr client.

The tests missed it because the one write test set VFBQUERY_CACHE_ENABLED=false to prove the guard, and that guard short-circuits before the URL is needed — the path that would have caught it was the path the test disabled.

Two tests that asserted content, not contract

term_info_serialization_dataset asserted exactly 4 types; DataSets gained stage labels and Ito2013 now has 5. test_default_excludes_are_a_strict_subset_of_everything demanded a strict decrease from excluding hemibrain and FAFB, which additionally requires those datasets to hold rows for that term pair — content, not contract. Both were failing on main.

Measured coverage

Population from each indexer's own get_parameters_query, coverage from SOLR:

anat_query                     51,848 expected   52,127 present    -279  (stale extras)
anat_image_query              635,685            632,476          3,209   0.5%
upstream_connectivity_query    12,960             12,671            289   2.2%
downstream_connectivity_query  12,960             12,694            266   2.1%
cluster_expression              3,290              3,290              0   0.0%

anat_query having 279 more documents than its population is stale entries for deprecated or removed classes. A fallback cannot fix that; only the bulk job clearing them can.

How to test

pytest src/test/test_term_info_fallback.py src/test/test_hierarchy.py — 54 pass, and the fallback tests need no backend. The example suite is unchanged at 15/15.

Scope note

The anat_query / anat_image_query rebuilds are wired but latent: the expansions those queries produce did not contain unindexed ids in any case I could construct, and get_instances / get_images_neurons return identically before and after. The demonstrated wins are the connectivity and hierarchy paths.