feat(dsl): lower a checked DSL program to the Scenario IR (p8-s1)

[JArmandoAnaya]

The first half of p8-s1: .osc in, ir::Scenario out. frontends/dsl/lower.h
turns a checked program into entities the existing engine accepts.

Lowering decides which DSL construct denotes which IR construct and nothing
else. Runtime semantics stay in the runtime, which both frontends share — the
architecture's rule, and the reason this is a small file.

Three decisions (ADR-0030)

The entry point is named, and a lone scenario names itself. §7.7.2 says
outright that selection "is defined by implementation", and that the latitude is
deliberate. LowerOptions::entry_point takes a qualified name (demo::overtake)
or the name as written. Empty means the root file's only scenario; a file
declaring several is an error that lists them. Guessing would make the run depend
on declaration order — exactly the hidden input the determinism contract exists
to remove. entry_points() returns the same list, in declaration order, because
that is what the file offers and a reader matches it against the file in front
of them.

A participant is a field deriving from std::physical_object. §8.7 roots its
actor hierarchy there. vehicle / person / stationary_object classify onto
ADR-0010's taxonomy; anything else deriving from physical_object stays an
unclassified participant. That last row is the interesting one: §8.7.10's
animal is a sibling actor, not a pedestrian category, and XML has nowhere to
put it. An identity and a control mode is all the runtime needs of it, and a
wrong classification would be worse than none.

Concrete means keep(field == constant). Either operand order, constant side
folding without a solver — that is what §6.3.1.2.1's "attribute-level concrete"
means, and it is the only shape whose meaning is unambiguous without search.
§7.3.8.2's conditional inheritance (inherits vehicle(vehicle_category == car))
is read the same way, since that is the spelling §8.7's own examples use.

Two things this deliberately does not do:

• It never converts. §7.3.4 folding already happened during checking, so a
  physical value arrives in its base unit. Re-applying §8.14.1.3's printed
  factors a second time is precisely the bug ADR-0029 exists to prevent.
• It never invents a number. §8.7 declares no performance limits at all —
  the domain model has no counterpart to XML's Performance. The IR's zeros are
  the faithful lowering, because the runtime already reads a non-positive limit
  as "unconstrained" (actor_max_speed). A category nothing fixes keeps
  ir::Vehicle's own default rather than a guess made here.

A pre-existing bug this surfaced

Constant evaluation resolved enum-name!member only in the current
namespace, never through the use list (§7.7.4.2). A scenario in
namespace demo use std writing vehicle_category!bus could not fold it — and
because the checker classifies constraints by whether they fold, every
enum-valued keep written that way was reported as needing a solver. It was
invisible until something had to read those values.

Fixed in expression.cpp with the ordinary lookup order (explicit ::, then the
current namespace, then the use list), and pinned in dsl_expression_test.cpp.
This is why the vehicle-category tests here are meaningful rather than vacuous.

Tests

dsl_lowering_test.cpp (17): entry-point selection in all four shapes, the
taxonomy mapping including the unclassified case, geometry and category binding
from both operand orders, a value written in cm arriving in metres, the
performance-is-unconstrained rule, the empty-scenario warning, and that lowering
the same source twice gives bit-identical geometry — load time is inside the
determinism contract.

dsl_expression_test.cpp (+1) for the enum-lookup fix.

1339 gtest (+18), 194 pytest.

What is left of p8-s1

The §8.8 movement-action subset, set_map_file → road backend, and scena-run
accepting .osc. Split off because this half stands on its own and is worth
reviewing on its own: it is where the mapping decisions live, and the action
subset is mechanical once they are settled.

Refs #44