fix(simulation): a loader is a callable returning the backend class

[cagataycali]

register_backend(name, loader) takes a zero-arg callable that returns the backend class. A class is itself callable, so passing one satisfied the registry's call and produced an instance where a class was expected: registration was accepted, and the failure surfaced later in create_simulation().

What

A SimEngine subclass passed as loader is refused at the door, naming the remedy (lambda: MyEngine, which also keeps the import deferred). force=True does not waive it - the mistake is not a name conflict. A class-shaped factory that is not a SimEngine stays accepted: that is a legitimate loader (row 5 above, byte-identical).

docs/api-reference.md spelled the parameter cls, so the documented call was the mistake: the keyword form raised TypeError: unexpected keyword argument 'cls', the positional form registered a class. The row now names loader.

Why the row survived

The reference's signature grader read only the leading code span on a table line. The drifted signature sat second in a cell that opened with list_backends(). It now grades every span on the line - 36 rows instead of 20, and that widening is what caught this one.

Tests

tests/simulation/test_factory.py pins the refusal, that force does not bypass it, that a refused call registers no alias, and three accepted-loader controls. Pre-fix 4 F / 47 P, post-fix 51 P. Gate: tests/simulation/ tests/*.py = 26,974 passed / 259 skipped / 0 failed; ruff + mypy clean.

LOC delta: +121 / -6 (+22 changelog fragment, +46 tests, +32 source, +25 grader).

[yinsong1986]

Summary

This PR closes a silent-failure hole in the simulation backend registry: register_backend(name, loader) expects a zero-arg callable returning the backend class, and because a class is itself callable, passing a SimEngine subclass directly was accepted at registration and only failed later inside create_simulation() (the registry calls the loader and reads __name__ off the result, so the caller got an AttributeError naming neither the parameter nor the registration that caused it). The fix refuses the class at the door with a TypeError naming the remedy (lambda: MyEngine), deliberately not waivable by force=True since the mistake is not a name conflict, while keeping the documented carve-out for a class-shaped factory that is not a SimEngine. The change also fixes the docs/api-reference.md row that spelled the parameter cls (the documented call was itself the mistake) and widens the reference's signature grader to read every code span on a table line rather than only the leading one -- which is what let the drifted second-position signature escape.

What's good

• The refusal converts a late, unattributed failure into a loud error with a remedy, and the narrowed input never worked -- so this is not a breaking change to any functioning caller.
• The new TypeError is added to the Raises: block in the same change (AGENTS.md > "A Raises: block names every refusal the function itself makes").
• Tests pin the refusal, that force does not bypass it, that a refused call registers no name or alias, and three accepted-loader controls including the deliberate non-SimEngine class-shaped-factory carve-out.
• The grader widening ships with a raised row floor (20 -> 30) and pins register_backend in the expected set, so the widening itself is graded rather than assumed; CI is green against the whole reference.
• Changelog fragment follows the <number>-<slug>.md convention; error strings are plain ASCII; scope is disciplined.