spacecost 0.3.1

Data contract unchanged at 1.15.0, no value moved, every delivery output
bit-identical to 0.3.0. An audit of what the module DERIVES against what it
merely states.

Four more lookups

The downleg departure burns were all typed at 0.3.0 on the argument that two
of six had no row. Four of them do, so four of them are lookups now:
cislunar off TLI -> NRHO insertion, mars_orbit off 1-sol Mars orbit -> Earth (TEI), mars_surface off the ascent-plus-TEI pair, and the 1,870 m/s
ascent component of lunar_surface off the descent row it is symmetric with.

Deriving what agrees and typing what does not is right; typing all six
because two could not be derived was not. That is the sharper version of the
0.3.0 note, and the correction is worth having: a blanket exception is how a
register stops being a decision.

🚨 TUG_ISP_S is 465 s and this package's hydrolox row says 452

The chains fly a 465 s cryogenic upper stage. PROPELLANTS_REFERENCE
carries hydrolox at 452 s, the RS-25 / RL-10 datasheet figure. 465 is the
top of the 450-465 s band an upper STAGE is quoted over, which is the right
figure for a tug and is what the consumer's published in-space prices were
computed with.

⚠️ Deriving it is a model change, not a refactor. Measured: swapping 465
for 452 moves the delivered price +2.96% cislunar, +3.29% geo, +3.63%
mars_orbit, +3.65% mars_surface, +5.21% lunar_surface. Those are published
downstream. Left as it is; the discrepancy is asserted at import instead,
so the row cannot move under the comment that describes it.

The same treatment is now applied to the GEO deorbit burn, which is 1,490 m/s
against a row of 1,488: typed, and pinned to the row it disagrees with.

A register, because a derivation claim decays one literal at a time

tests/test_delivery.py reads the module's source and requires every
numeric literal to be either algebra (0 and 1) or a row on TYPED saying
which table cannot supply it. No test of the outputs can catch this: a
hardcoded 3,600 and a looked-up 3,600 give identical numbers and identical
hashes right up until the row moves and only one follows.

⚠️ Both halves are findings. A literal with no row is a value that
stopped being derived; a row with no literal is a permission still granted for
a number that has gone. Proved by planting one of each, and by planting a
literal that changes no value at all -- the case the import assertions cannot
reach.