### CASE 2: one event, 27 boards
27 deals, best of 5 rounds

threads        ours        dds        ref   (deals/sec)
1              10.2       12.4       10.4   vs dds: ours 0.82x, ref 0.84x
4              37.1       46.6       35.5   vs dds: ours 0.80x, ref 0.76x
8              70.8       90.1       59.3   vs dds: ours 0.79x, ref 0.66x
12             77.6       98.4       68.2   vs dds: ours 0.79x, ref 0.69x

Re-measured after the (deal, strain) work split. The same machine, same
session, immediately before, with whole deals as the work item:

threads        ours        dds        ref   (deals/sec)
1              10.2       12.4       10.4   vs dds: ours 0.82x, ref 0.84x
4              34.6       46.7       35.5   vs dds: ours 0.74x, ref 0.76x
8              59.2       91.0       59.2   vs dds: ours 0.65x, ref 0.65x
12             60.0       99.8       68.0   vs dds: ours 0.60x, ref 0.68x

Two interleaved passes of old/new agreed to within 2% on every cell, which is
what makes the difference above readable: the machine is not still enough for a
single pass of each to be worth quoting.

### CASE 3: bulk, 500 deals
500 deals, best of 3 rounds

threads        ours        dds        ref   (deals/sec)
1               8.8        9.3        9.2   vs dds: ours 0.94x, ref 0.99x
4              31.8       34.7       35.0   vs dds: ours 0.92x, ref 1.01x
8              60.6       61.3       64.3   vs dds: ours 0.99x, ref 1.05x
12             71.0       67.3       78.3   vs dds: ours 1.06x, ref 1.16x

The same run with whole deals as the work item, for the regression check:

threads        ours        dds        ref   (deals/sec)
1               8.8        9.3        9.2   vs dds: ours 0.94x, ref 0.99x
4              32.3       34.9       35.0   vs dds: ours 0.93x, ref 1.00x
8              59.6       60.6       63.7   vs dds: ours 0.98x, ref 1.05x
12             72.0       64.9       77.0   vs dds: ours 1.11x, ref 1.19x

Unchanged: -1.5%, +1.7%, -1.4% at 4, 8 and 12 threads, in no consistent
direction and inside the 1-2% these numbers wander anyway. With 2,500 work
items either way there was never any imbalance here for a finer unit to fix.
