spacecost 0.3.2

A test fix, and it is a fix to the CONTRACT rather than to a number. No
package code changed, no value moved, the data contract stays 1.15.0.

🚨 0.3.1's tests claimed a cross-platform guarantee this package declines to make

tests/test_delivery.py asserted == on every platform. CI went red on the
first push, on lunar_surface and nothing else: Linux returns
21209.958393766807 where Windows returns 21209.9583937668. One ULP.

That was not a defect in delivery.py. It was the wrong contract, and this
package already had the right one written down. The output is a
contract says the six reference tables
are byte identical on every platform BECAUSE they pass values through, and the
composite summary is promised only "the same values, to a few ULP" BECAUSE it
runs through exp() -- the platform libm, which IEEE 754 does not require to
be correctly rounded. Every number in delivery.py is a mass ratio, so
every number in it is exp(). The tests were written to the consumer's
bit-identity standard and imported it into a package that says, in its own
README, that it cannot promise that here.

⚠️ lunar_surface failing ALONE is the tell. It is the only chain with
two burns at different dry-mass fractions, so it compounds the most rounding;
the single-burn chains happened to land on the same float. A cross-platform
claim that holds for six of seven cases is a claim that has not been tested.

✅ Fixed with test_parity.py's own pattern rather than a new one: exact
on the platform reference/summary_meta.json was recorded on, and a 1e-12
relative tolerance elsewhere. Not a relaxation -- on the recording platform it
is the same exact comparison, and the tolerance only applies where an exact
one would assert something exp() cannot deliver.

⚠️ The tolerance was chosen against the smallest REAL change this module
could suffer, not by eye. Deriving TUG_ISP_S is the smallest at 2.96%;
1e-12 sits nine orders of magnitude below that and four above the observed
libm spread. Proved by feeding the checker the actual value CI returned (it
passes), the value the Isp swap would give (it fails), a 1e-9 drift (fails)
and a 1e-15 one (passes) -- and by confirming the exact comparison still bites
on the reference platform.

⚠️ economicspace is unaffected and does not have to move for this. It
runs on the reference platform, its own bit-identity is unchanged and
independently verified, and it does not run this package's test suite. The
repin to 0.3.2 is so the tag it pins has a green suite, not because anything
it computes changed.