fix(recall): widen context tag candidate fetch

[jack-arturo]

Summary

• widen semantic vector candidate fetches when recall has context_profile.priority_tags, so context_tags boosts can promote candidates outside a small requested limit
• add lab harness support for per-query context_tags
• add a small context-tags probe set for issue recall: context_tags boost dropped at small limit (candidate fetch only widens for hard tag_filters) #201

Closes #201.

Public/API Notes

• /recall response shape is unchanged.
• context_tags remains a soft boost; this only widens the pre-scoring vector candidate pool when a semantic query or embedding is present.
• Lab test-set rows may now include context_tags.

Validation

• RED confirmed before implementation:
  • tests/test_recall_overfetch.py::test_context_tags_widen_vector_fetch_before_scoring failed with requested vector limit 20 instead of 50
  • tests/lab/test_lab_corpus.py::test_recall_serializes_params failed with unexpected context_tags
  • tests/lab/test_run_recall_test.py::test_run_single_query_forwards_row_context_tags failed because row-level tags were ignored
• PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 .venv/bin/pytest tests/test_recall_overfetch.py tests/test_context_tag_separator.py -q -> 17 passed
• PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 .venv/bin/pytest tests/lab/test_lab_corpus.py tests/lab/test_run_recall_test.py -q -> 11 passed
• PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 .venv/bin/pytest tests/test_api_endpoints.py -q -k recall -> 69 passed, 124 deselected
• make test -> 631 passed, 1 skipped, 65 deselected
• make lint -> passed
• git diff --check -> passed
• python -m json.tool lab/test_sets/queries_context_tags_20260704.json -> passed
• make bench-eval BENCH=locomo-mini -> 84.26% overall accuracy (198/235), 367.5s, 69 cat-5 questions skipped by judge-off run; matches latest comparable local LoCoMo-mini artifact
• make bench-eval BENCH=locomo -> 81.91% overall accuracy (1263/1542), 2072.6s, 444 cat-5 questions skipped by judge-off run; slightly above latest comparable local LoCoMo-full artifact (81.84%, 1262/1542)

Live Lab A/B

Ran two isolated local stacks from the same restored production snapshot, prod-api-20260611-post-repair-001241:

• baseline: origin/main (7636021) at http://localhost:8011
• candidate: this PR (a29e9fb) at http://localhost:8012
• both reported memory_count=9900, vector_count=9900, sync_status=synced

Context-tag probes (lab/test_sets/queries_context_tags_20260704.json, limit=5, current_only=false):

• baseline: Recall@5/10/20 0.000, MRR 0.000, NDCG@10 0.000, 5 misses
• candidate: Recall@5/10/20 1.000, MRR 1.000, NDCG@10 1.000, 0 misses

Plain production set (lab/test_sets/queries_prod_20260611.json, limit=20, current_only=true):

• baseline: Recall@10 0.045, MRR 0.024, NDCG@10 0.029, latency avg 765ms
• candidate: Recall@10 0.275, MRR 0.220, NDCG@10 0.231, latency avg 972ms
• observed no regression; latency increased by about 208ms avg on this local restored stack

The context-tag probe is the direct #201 validation signal. The plain-query production lift is observational and likely reflects existing derived context_profile.priority_tags now widening the candidate pool as intended. Ignored result files are under lab/results/issue201/, including issue201_live_ab_summary.json.