perf(summary): use dataset-record fields when /document-class-counts times out

[audriB]

Why

Smoke-test pass after PRs #101/#102/#103 found that for the largest production datasets (Jess Haley 78k docs, Sophie Griswold 101k docs) the `/document-class-counts` endpoint still exhausts its 20s stage-1 deadline on every synthesis attempt. The cloud's Mongo aggregation on `className` without proper indexes simply doesn't fit in the budget.

Net effect: every cron warm cycle landed a degraded summary with all zero counts AND null per-class facts (because `counts.subjects=0` short-circuited stage 2's openminds_subject fanout). The DatasetSummaryCard rendered "0 sessions, 0 subjects, …, Not applicable" everywhere, even though the dataset record itself reports `numberOfSubjects=1656` and `documentCount=78687`.

The frontend band-aid (Waltham-Data-Science/ndi-cloud-app#91) papers over this on the rendering side by enriching the degraded summary with record fields. This PR is the real fix at the backend: when stage-1 counts times out, synthesize a counts envelope from the dataset record's raw fields and gate stage 2 from those fields directly. Stage 2 still attempts per-class fanouts; the species/brainRegions/strains/sexes ndiqueries succeed in isolation (only `/document-class-counts` is pathologically slow on these datasets).

Before vs after

```
Before:
counts: { sessions: 0, subjects: 0, probes: 0, elements: 0,
epochs: 0, totalDocuments: 0 }
species: null ← stage 2 short-circuited
brainRegions: null ← stage 2 short-circuited
extractionWarnings: ["class counts query failed: ..."]

After (Jess Haley):
counts: { sessions: 0, subjects: 1656, probes: 0, elements: 0,
epochs: 0, totalDocuments: 78687 }
species: [Caenorhabditis elegans, Escherichia coli] ← REAL stage-2 data
brainRegions: [whole nervous system] ← REAL stage-2 data
strains/sexes: real data when available
extractionWarnings: ["class counts query failed: ..."]
```

Worst-case timing budget

Stage	Time
stage 1 (counts + dataset metadata, parallel)	20s (counts hits deadline; metadata succeeds fast)
stage 2 (3 per-class ndiqueries in parallel)	25s (each bounded by per-class deadline)
ontology resolution	~2s
total	~47s (well under Railway's 88s ceiling)

Caveat

Per-class counts (sessions/probes/elements/epochs) stay 0 when stage-1 counts times out — the dataset record doesn't expose these. They display as 0 with the "X warnings" tooltip explaining the underlying cause. A future iteration could compute these from `/dataset/:id/documents?class=...&pageSize=1` envelopes (4 extra calls per degraded synthesis) — not worth it for fields the user rarely notices vs the now-restored species/brainRegions facts.

Coverage

• `test_stage_1_counts_timeout_still_runs_stage_2_via_record_fields`: end-to-end pin asserting stage 2 attempts and succeeds despite counts timeout
• `test_safe_record_int_handles_all_input_shapes`: defensive helper accepts any input shape and degrades to 0

Test plan

• `pytest backend/tests`: 557 passed, 1 skipped (opentelemetry)
• `ruff check backend/`: clean
• `mypy backend/`: 56 source files, no issues
• Smoke-test on Railway post-deploy: confirm Jess Haley summary returns species + brainRegions populated despite counts-timeout warning

🤖 Generated with Claude Code