[neomatrix369]

What and why

Slice 67E, backend half. The stats tile for "Query Results" shows a 6-7 digit number with no way to see what it is made of. This adds the backend facts needed to break it down by golden-dataset record; the tile that displays it is the frontend half (separate branch, after this).

• Each stored result row is one query answered in one run, so total_results is the sum over runs of queries answered (runs × golden queries when every run finished). top_k only changes how many chunks each row holds. unique_queries is not the golden count: it equals total_results because a fresh id is generated for every query in every run.
• New per-experiment stats: result_runs, golden_queries, golden_queries_source (config or max_per_run), partial_runs. New group totals: total_result_runs, golden_queries_min, golden_queries_max. All are assembled once in server/db/ports/stats_common.py (golden_query_facts) for Mongo, Postgres and SQLite; adapters only supply per-run row counts they already build, so the single-experiment paths issue no extra query.
• The golden count is the number of records in the golden dataset, counted with the pipeline's own load_queries and saved as sweep_summary.golden_queries the first time a run loads the queries (once per process, best effort, idempotent). Older experiments are inferred from stored results and tagged max_per_run.
• Redis and Elasticsearch report nothing new in the grouped view (no overlay, as decided). The detail view overlays the run-state facts for every backend.
• Failure isolation: an optional per-run query that fails logs one warning and reports unknown (None), never zero, and never fails the endpoint.

Verification

• ./scripts/ci/quality-gates.sh: all gates passed (1015 backend tests, coverage 79.31 / 66.48 / 79.31 against the 72/59/72 floors, ruff, mypy, xenon, bandit, audits).
• Seam test drives the real _run_single against a real SQLite store with a 3-record, two-persona, duplicate-text queries file and reads sweep_summary.golden_queries == 3 and exactly 3 result rows back; deleting the persist call or passing a wrong length makes it fail.
• Live runs against throwaway containers (the live suites are skipped in PR CI): Postgres integration tests 20 passed; new live Mongo test 2 passed (two-stage grouped pipeline, single and grouped paths agree).
• Three independent reviews (code quality, system design, acceptance coverage) approved after one round of changes: failure isolation and duplicate aggregates, an unwitnessed orchestrator seam, a Postgres test that had never executed, and a missing import in an uncommitted edit.

Notes for the reviewer

• Known limits: under the default split-store setup the grouped view has no run-state overlay, so the group golden facts stay empty (the same pre-existing gap total_results has); partial_runs also counts runs still in flight; the stored count is write-once and goes stale if queries_file changes between pause and resume; the Mongo and Postgres live tests do not run in PR CI.
• Follow-ups (listed in the spec): an atomic set-field operation for sweep_summary (the SQLite read-modify-write can lose a concurrent update; a lost count self-heals through the fallback); merge the two Postgres grouped scans; the Mongo per-experiment query_ids document-size ceiling; a Mongo index on results(experiment_id, run_id).
• Docs: spec, PROGRESS, CHANGELOG, GROWTH, DECISIONS #313. This PR and fix(dashboard): numbers that add up: one Bayesian denominator, pure label helpers (slice 67D, first bundle) #269 (67D) both add lines at the same spots in those files; whichever merges second needs a trivial rebase.
• No attribution trailers.

Complexity

Automatically refreshed by CI for product-code changes.

Scope	Tool	Result
Python	Radon	highest CC 40 (rank E), 900 blocks
JS/TS	ESLint	18 function(s) above CC 10
JS/TS	ESLint sonarjs	7 function(s) above cognitive 15

Functions above the JavaScript threshold

• ExperimentDetailScreen.tsx:53 — CC 118
• SearchExplorerScreen.tsx:35 — CC 38
• ExplorePanels.tsx:415 — CC 29
• LoadingFeedbackPanel.tsx:41 — CC 28
• labels.ts:36 — CC 25
• ExperimentsScreen.tsx:64 — CC 21
• VectorDbStatsPanel.tsx:84 — CC 18
• ExperimentControlButtons.tsx:54 — CC 14
• loadWalkthruResults.ts:37 — CC 14
• ExperimentsScreen.tsx:465 — CC 13

Functions above the cognitive-complexity threshold

• ExperimentDetailScreen.tsx:53 — cognitive 60
• LoadingFeedbackPanel.tsx:41 — cognitive 27
• labels.ts:36 — cognitive 25
• ExplorePanels.tsx:415 — cognitive 23
• markdown.ts:59 — cognitive 20
• SearchExplorerScreen.tsx:35 — cognitive 17
• steps.ts:63 — cognitive 16