test(compat): vary lot-identity axes together in the booking fuzzer - #2117

The first campaign answered one question well and never asked the
others. Every divergence it found had same-date lots (12/12, against
0/189 for seeds without them), because random generation collided on
date, cost and label only by accident: two lots at the same HIGHEST cost
under HIFO, or two on the same date under STRICT, essentially never
appeared.

Lot identity has three components a method can tie on, so the generator
now constructs each configuration deliberately — none, date, cost, both,
label — and reports the tie mode with every divergence so the classes
stay separable. All 20 method x tie cells are populated (17-31 runs per
cell at --runs 500), which is what makes the empty ones meaningful:
HIFO with tied costs and STRICT with tied dates now have real coverage
and agree with beancount.

Two corrections during the extension, both caught by measuring rather
than reading:

`none` was not a control. Costs were drawn with replacement from a
four-element pool, so 79% of "no tie" runs quietly contained a cost tie
and could not be used to attribute a divergence to the date axis. It now
samples without replacement: 0 date ties, 0 cost ties.

Tying exactly two lots LOST the FIFO coalescing divergence the first
campaign had found. That shape needs three lots sharing a date with two
of them sharing a cost — a NON-CONTIGUOUS repeat — and a uniform draw
produced it about once in seven date-tied runs. Ties now span a
random-sized group, and a date-tied group draws its costs from a
deliberately narrow pool so repeats land non-contiguously. Verified
against the known-divergent ledger directly: the bug was still there,
the generator had stopped reaching it.

At --runs 500 this finds 10 divergences: 8 LIFO x date, 1 FIFO x date,
and one new class — LIFO x label, where two lots differ only by label
and rledger REJECTS a ledger beancount books. That is the mirror of the
acceptance divergence already reported, so the disagreement cuts both
ways. All are the same root cause and are tracked in #2115.

Both from review, and both were real.

The generator gained a label axis but the comparison key did not, so two
lots differing only by label collapsed into one entry on BOTH sides. The
fuzzer could see a label divergence only when it changed acceptance; a
divergence in which labelled lot got consumed was invisible by
construction. Adding `cost_label` to the key took the campaign from
490/500 to 465/500 — 15 LIFO x label divergences where there had been 1.
Seed 30 is representative: same total, same cost, same date, but
rledger books lot-a=5/lot-b=1 where beancount books 3/3.

The narrowed cost pool was described as producing non-contiguous repeats
and did not. It drew independently per lot, so a group of two never
repeated and a group of three usually came out contiguous or all-equal.
The shape is now constructed outright ([a, b, a, ...]): 47 of 47
date-tied groups of three or more have it, against roughly one in seven
before. FIFO divergences went from 1 to 7.

Net effect of a comment that only claimed something and a key that only
covered part of the identity it was generating: two thirds of the
divergences in this space were not being reported.