feat: units and geometry - #2

…of a corner

Layer 1 of RFC-005's stack, carrying refactors R1, R2 and R5's half. Two small
modules with no OpenCV dependency, so this layer runs anywhere -- the imaging
half starts at layer 2.

R2 -- the measured constants get one home, and it fixes a live inconsistency.
Three modules each carried their own copy of the card geometry and two of them
disagreed: `detect.CardSpec.aspect` defaulted to 0.5843, which has no provenance
in any document, while `project.DEFAULT_ASPECT` used the calipered 0.583 and
`cli._settings` separately hardcoded 70.0 x 120.0. The consequence was live --
`deckle detect` outside a project gated against a different card than the same
command inside one. `units.py` already owned DEFAULT_DPI on exactly this
reasoning, so DEFAULT_CARD_MM and DEFAULT_ASPECT join it and the copies are
removed as layers 4, 7 and 10 introduce the modules that held them.

Measured on the three padded scans before choosing: cards report aspect 0.5838,
which is 0.0008 from 0.583 and 0.0005 from 0.5843, against a tolerance of 0.010.
Both pass with more than ten times the margin, all three scans detect 4/4 under
either, and the reported geometry is byte-identical -- aspect feeds the gate, not
the fit. So this is a provenance fix, not a behaviour change, and it is worth
saying which it is.

R5 -- DEFAULT_STRATEGY moves here too, which cuts the edge that ran the wrong way
in the import graph: `project.py` imported it from `edges.py` purely so `Config`
could default it, making the state-file module depend on the edge fitter. The
STRATEGIES tuple comes along, because which strategy a scan needs is a fact about
how the scan was taken -- with the foam pad the true edge is a hard 150-248 luma
step, without it there is a shadow ramp in the clearance gap -- and that is rig
description, not fitter internals. Keeping the pair together also makes
"the default is one that exists" testable here rather than at layer 3.

R1 -- `geometry.corners_from_lines()`. `jig._corners_from_walls` and the inline
block in `detect.detect_card` built the identical TL/TR/BR/BL intersection of four
fitted lines. RFC-001 defines a corner as that intersection, because a real corner
is rounded and there is nothing there to observe, so one place should say so. The
winding is documented as a contract rather than an implementation detail: `rectify`
maps the quad onto the output rectangle in that order.

Tests: the corner primitive against a clean rectangle, against the 1.2deg of skew
hand placement produces, and against edge stubs that stop well short of the
corners -- that last one is the property the primitive exists for. Plus the
constants against the calipers, and a check that 0.5789, the deck spec default,
is not what we use.

`just test` 31 passed, `just lint` clean, standalone at this layer.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>