feat: consume typed OperatorDescriptor from op_system in flepimop2-op_engine adapter

[jc-macdonald]

Motivation

Companion to ACCIDDA/op_system#156.

Currently flepimop2-op_engine retrieves operator specs from system.option("operators", None) as an untyped dict and has to reverse-engineer the model axis structure before calling build_crank_nicolson_operator. There is no enforced contract between what op_system produces and what this adapter expects.

Proposal

Once ACCIDDA/op_system#156 lands a typed OperatorDescriptor dataclass in op_system's public API, update the flepimop2-op_engine adapter to:

1. Read system.option("operators") as tuple[OperatorDescriptor, ...] rather than an opaque dict.
2. Build CN (L, R) matrices from the typed descriptor fields:

   for desc in system.option("operators", ()):
       L, R = build_crank_nicolson_operator(desc.geometry, desc.coefficient, dt)

3. Drop any internal parsing/coercion logic (_coerce_operator_specs) that currently normalises the untyped option.

What stays in op_engine

• build_crank_nicolson_operator(geom, cfg, dt) — CN matrix construction at solve time (depends on dt, which is solver-side)
• StageOperatorFactory and per-step factory calls for adaptive / TR-BDF2 schemes
• LU factorisation caching in implicit_solve

All of this remains entirely correct. Only the source of the operator geometry changes from an untyped option to a typed descriptor.

Scope

• Update OpEngineFlepimop2Engine in flepimop2-op_engine/src/flepimop2/engine/op_engine/__init__.py
• Remove or simplify _coerce_operator_specs
• Add a type annotation / isinstance guard on the incoming option value
• Update tests in flepimop2-op_engine to pass OperatorDescriptor instances

Dependency

Blocked on ACCIDDA/op_system#156.

[jc-macdonald]

Update: op_system#156 (which this issue was blocked on) is now closed — OperatorDescriptor shipped, but with a different field shape than this issue's code sample assumed. The actual shape (op_system/src/op_system/_operators.py):

@dataclass(frozen=True, slots=True)
class OperatorDescriptor:
    axis: str
    kind: str | None = None       # e.g. "advection", "diffusion", "transport", "jump_integral"
    bc: str | None = None         # e.g. "absorbing", "periodic", "neumann", "reflecting"
    velocity: str | None = None   # parameter *name* for advection velocity
    rate: str | None = None       # parameter *name* for diffusion rate/coefficient
    kernel: Mapping[str, Any] | None = None

There is no desc.geometry/desc.coefficient — velocity/rate are parameter names to resolve from params/config, not resolved geometry or coefficient values (grid geometry stays an op_engine-side concern, consistent with this issue's original "What stays in op_engine" section). This issue's scope (drop _coerce_operator_specs, read system.option("operators") as tuple[OperatorDescriptor, ...]) is still valid — just update the code sample to build CN matrices from desc.axis/desc.kind/desc.bc plus the resolved velocity/rate parameter values pulled from params, not from descriptor fields that don't exist.