mvp_report: ④ accuracy breaks out per sketch family (5 sub-rows + raw)

[zzylol]

Summary

Refs #46. Renders ④ accuracy as 6 rows (5 sketch families + raw passthrough) instead of one aggregate, with per-family ε bounds and a single PASS/FAIL/UNKNOWN rollup in §2.

Per-family bounds (hardcoded; matches the shared MVP-demo contract):

Family	Metric	Query class	Bound
DDSketch	http_latency_ms	quantile	rel_err ≤ 0.01
KLL	request_size_bytes	quantile (rank-err)	rel_err ≤ 0.005
HLL	unique_users_per_min	cardinality	rel_err ≤ 0.0325
CountSketch	top_endpoint_qps	top-K	recall ≥ 0.85
CountMinSketch	endpoint_request_freq	frequency	rel_err ≤ 0.02 (≈ 1/w, w=64)
raw	http_requests_total	sum_rate	exact (rel_err == 0)

Rendered ④ table shape

## §3 Accuracy (per sketch family)

| Sketch family | Metric | Query class | n | rel-err / recall | within ε bound? | Verdict |
|---|---|---|---|---|---|---|
| DDSketch | http_latency_ms | quantile | 300 | 0.0042 | yes (≤0.01) | PASS |
| KLL | request_size_bytes | quantile | 300 | 0.0030 (rank-err proxy: reducer reports value rel-err) | yes (≤0.005) | PASS |
| HLL | unique_users_per_min | cardinality | 300 | 0.0180 | yes (≤0.0325) | PASS |
| CountSketch | top_endpoint_qps | top-K | 300 | recall=0.920 | yes (≥0.85) | PASS |
| CountMinSketch | endpoint_request_freq | frequency | 300 | 0.0140 (additive proxy vs 1/w ≈ 0.02 ceiling) | yes (≤theoretical (1/w)) | PASS |
| raw | http_requests_total | sum_rate | 300 | 0.0000 | yes (exact (=0)) | PASS |

**Verdict ④:** PASS  · 6/6 families within bound; FAIL=0 UNKNOWN=0

Aggregate-verdict logic

The §2 row for criterion 4 reads Accuracy (per sketch family, see §3) and shows the rolled-up verdict:

• PASS iff every family's observed value is within its bound AND raw is exact (== 0)
• UNKNOWN if any family has no rows (bubbles up)
• FAIL if any family is over-bound (bubbles up — UNKNOWN dominates over FAIL)

Detail summary in the §2 detail column: 6/6 families within bound; FAIL=0 UNKNOWN=0.

Honest caveats

• KLL's reducer (accuracy_reduce.py) emits rel_err as value-relative-error, not rank-error. We compare it against the same 0.005 threshold and annotate the row as a "rank-err proxy". Fixing this requires the reducer learn to emit rank-err — out of scope for this PR.
• CountMinSketch's additive-error bound (1/w against sum-of-frequencies) doesn't have a dedicated reducer column either; we reuse rel_err against a hardcoded 0.02 ceiling (≈ 1/w with w=64). Same fix path: extend the reducer.

Test plan

• All 14 pre-existing test_mvp_report.py tests pass unchanged (one assertion updated to match the renamed §2 row).
• 8 new tests added covering: all-pass, single-family fail, missing-family UNKNOWN, raw-must-be-exact, empty-csv UNKNOWN, table shape, single-mode end-to-end, dual-mode end-to-end.
• Hand-rendered against a synthetic fixture covering all 5 sketch families + raw — output matches the contract.
• Once a real run produces an accuracy.csv with all 6 families, verify the §3 table renders without manual fixture intervention.

🤖 Generated with Claude Code