Shape matchers: recognizing families in a parsed equation

[roed-math]

The shape matchers: matchers.run(parsed) returns Match(slug, data) records.

This is the file where a human eye is most useful on the mathematics: the
degree/genus tests, the reductions to standard forms, and the data each match
extracts (D for Pell, a-invariants for Weierstrass, the regime of a
generalized Fermat equation, ...). Matchers never mutate the parsed equation.

API note: Match.transform is a CoordinateTransform (transforms.py),
not a prose string. It holds both directions of the map between the user's
variables and the family's standard coordinates, push_forward/pull_back,
and the structural roles a family assigns — which have moved out of
Match.data["roles"], since a solver needs them together with the map that
consumes them. A string cannot transport a solution: 5*x^2 - y^2 = 1 is a
Pell equation only after reading the user's y as the standard x, and the
answer has to come back. Operations that change the equation but no
coordinate (multiplying through by -1) are recorded separately, in
.operations.

Two recognition fixes worth a look: a parameter in a coefficient is no longer
dropped (a Term keeps its parameter factors apart from coeff, so A*2^n
and 2^n shared coeff == 1 and Pillai matched both), and orienting
D*x^2 - y^2 = N as a Pell equation swaps the variables, which negates the
right-hand side too — without that, 5*x^2 - y^2 = 1 was matched as the
different equation x^2 - 5*y^2 = 1.

The recognizers cover 44 families, most of which are not registered yet. That
is deliberate and safe: a recognizer whose family has not landed is inert,
because the classifier (next PR) ranks matches through the registry and drops
slugs it does not know. tests/test_registry.py checks the flags of the
registered ones; 99-polish tightens that to every emitted slug once the
registry is complete.

Part of the series that splits #1 into reviewable pieces:

1. 01-bibliography — packaging, docs, annotated bibliography
2. 02-parsing — equation strings to a term model
3. 03-registry — the YAML family registry (3 seed families)
4. 04-matchers — shape recognizers
5. 05-classify — the classification pipeline
6. 06-solvers — solver framework, two seed solvers, and the CLI
7. 07..09-backbone — the 23 parent families of the DAG, by depth
8. one PR per remaining family (38 of them, mutually independent)
9. 99-polish — restore the full doctests and tighten the invariants