Update benchmark numbers from latest 10M run, prune invalidated 100M results

Incorporate new4 benchmark output (commit a094df7): honest 1T window
numbers, corrected CB-CD-GRP, improved LIKE/string/CORR ratios.
Remove 12 invalidated 100M rows (WIN, LIKE, H2O-Q6/Q8/Q9, DS-Q98, Q27).
35 of 46 wins at 10M, 15 of 34 at 100M.

Author: whilo

README.md

@@ -49,38 +49,38 @@ Stratum vs DuckDB (10M rows, single-threaded, best of array/index mode)
 ══════════════════════════════════════════════════════════════════════════
                                     Stratum    DuckDB    Ratio
 ──────────────────────────────────────────────────────────────────────
-TPC-H Q6 (filter+sum-product)        15ms       31ms    2.1x faster
-SSB Q1.1 (filter+sum-product)        14ms       30ms    2.1x faster
-Filtered COUNT (NEQ pred)              4ms       13ms    3.6x faster
-Group-by COUNT (3 groups)             19ms       27ms    1.4x faster
-TPC-H Q1 (7 aggs, 4 groups)          79ms       98ms    1.2x faster
-H2O Q3 (100K string groups)          70ms      380ms    5.4x faster
-H2O Q2 (10K string groups)           32ms      126ms    3.9x faster
-H2O Q6 (STDDEV, 10K groups)          31ms       83ms    2.7x faster
-H2O Q9 (CORR, 10K groups)            68ms      136ms    2.0x faster
-H2O Q10 (10M groups, 6 cols)        842ms     7294ms    8.7x faster
-LIKE '%search%' (string scan)         54ms      252ms    4.7x faster
-AVG(LENGTH(URL)) (string fn)          40ms      179ms    4.5x faster
+TPC-H Q6 (filter+sum-product)        13ms       28ms    2.2x faster
+SSB Q1.1 (filter+sum-product)        13ms       28ms    2.2x faster
+Filtered COUNT (NEQ pred)              3ms       12ms    4.0x faster
+Group-by COUNT (3 groups)             17ms       24ms    1.4x faster
+TPC-H Q1 (7 aggs, 4 groups)          75ms       93ms    1.2x faster
+H2O Q3 (100K string groups)          71ms      362ms    5.1x faster
+H2O Q2 (10K string groups)           33ms      123ms    3.7x faster
+H2O Q6 (STDDEV, 10K groups)          30ms       81ms    2.7x faster
+H2O Q9 (CORR, 10K groups)            61ms      134ms    2.2x faster
+H2O Q10 (10M groups, 6 cols)        832ms     7056ms    8.5x faster
+LIKE '%search%' (string scan)         47ms      240ms    5.1x faster
+AVG(LENGTH(URL)) (string fn)          38ms      170ms    4.5x faster
 ──────────────────────────────────────────────────────────────────────
-COUNT(DISTINCT) 1.1M                 116ms       62ms    DuckDB 1.9x
-COUNT WHERE sparse pred               21ms        5ms    DuckDB 4.2x
+COUNT(DISTINCT) 1.1M                 109ms       61ms    DuckDB 1.8x
+COUNT WHERE sparse pred               22ms        5ms    DuckDB 4.9x
 ──────────────────────────────────────────────────────────────────────
 ```
 
 **Summary (10M rows, single-threaded, best of array/index mode vs DuckDB 1T):**
 
 | Category | Queries | Stratum faster | DuckDB faster |
 |----------|---------|---------------|--------------|
-| TPC-H / SSB | 7 | 7 (1.2–3.6x) | 0 |
-| H2O Group-By | 10 | 10 (1.0–8.7x) | 0 |
+| TPC-H / SSB | 7 | 7 (1.2–4.0x) | 0 |
+| H2O Group-By | 10 | 9 (1.0–8.5x) | 1 |
 | H2O Join | 3 | 2 (1.1x) | 1 |
-| ClickBench | 22 | 16 (1.0–119x) | 6 |
-| NYC Taxi | 4 | 2 (1.1–1.4x) | 2 |
+| ClickBench | 22 | 15 (1.0–105x) | 7 |
+| NYC Taxi | 4 | 2 (1.0–1.4x) | 2 |
 | Statistical | 5 | 5 (1.1–2.4x) | 0 |
-| Window | 3 | 3 (2.4–3.2x) | 0 |
-| Hash Join + TPC-DS | 3 | 2 (1.9–2.9x) | 1 |
+| Window | 3 | 3 (1.3–2.1x) | 0 |
+| Hash Join + TPC-DS | 3 | 2 (1.3–2.1x) | 1 |
 
-DuckDB wins on sparse-selectivity filters and global COUNT(DISTINCT). Stratum wins on arithmetic-heavy aggregates, string operations, group-by, LIKE, statistical aggregates, and window functions.
+DuckDB wins on sparse-selectivity filters, global COUNT(DISTINCT), and window-based top-N. Stratum wins on arithmetic-heavy aggregates, string operations, group-by, LIKE, statistical aggregates, and window functions.
 
 Run benchmarks yourself: