benchmark: q8 ground truth

[GnaneshGnani]

Q8 — Price anomaly IQR fence (DDSketch precision/recall/F1)

Replaces the frac_iqr_lt_10pct metric with a Tukey fence anomaly detection evaluation: precision, recall, and F1 score for flagging stocks with price outliers. Sets accuracy_sla=0.0001 for Q8 (tighter DDSketch bucket width).

Benchmark results

Setup: Day 08-11-21, data_filtered, accuracy_sla=0.001, 4 × 15-min windows
Sketch: DDSketch [0.25, 0.75]; Tukey fence k=1.5
Thresholds: precision ≥ 0.70, recall ≥ 0.80, F1 ≥ 0.75

Window (CET)	Valid IQR symbols	Precision	Recall	F1
09:00–09:15	37.7%	0.776	0.450	0.570
09:15–09:30	37.3%	0.731	0.362	0.484
09:30–09:45	34.2%	0.653	0.278	0.390
09:45–10:00	30.3%	0.547	0.257	0.349
Aggregate	—	0.719	0.368	0.487

❌ Fails — all windows fail recall threshold. Root cause: at accuracy_sla=0.001, ≈60–70% of symbols collapse to sketch_iqr=0 and are excluded; remaining symbols have overestimated IQR (2–3×), making the sketch fence too wide → borderline anomalies missed → low recall.

Next step: Re-run with accuracy_sla=0.0001 (already applied in run.py).

Cumulative: Q1: 11/12 ⚠️ | Q3: not run | Q4: 12/12 ✅ | Q5: 3/12 ⚠️ | Q6: 12/12 ✅ | Q7: 12/12 ✅ | Q8: 0/4 ❌

See results/benchmarking.md for multiplier sweep and full analysis.

[zzylol]

Closing — this branch is conflicting with main (badly diverged, ~32k-line deletions) and predates the consolidated multi-sketch + latency eval harness merged in #494, which redid this ground-truth work. Please reopen and rebase if any ground truth here is still needed.