_Trying to address the elephant in the room._

## Parse-time evaluation

Different levels:

0. No parse-time eval
1. Can evaluate simple constants (Done)
* Example: `const name = 'spam.nu'; use $name`
2. Able to evaluate simple calls (https://github.com/nushell/nushell/pull/9499)
* Example: `const NU_LIB_DIRS = [( '~/.nuun/overlays' | path expand)]`
* Only "simple" commands are allowed
* _Seems_ to work fine but I can't prove it (see 4)
3. Able to evaluate complex calls that reference the working set
* Example: `const x = (if true { 'a' } else { 'b' })`
* This panics because the blocks are in working set, not engine state, during parsing
* Requires API change: `Command::run()` would need to take `struct EngineSetView { engine_state: &EngineState, delta: &StateDelta }`
----
4. Able to evaluate calls with parser side effects (parser keywords)
* `Command::run()` would need to take a mutable reference to working set
5. 1:1 parity between runtime and parse-time eval

I'd draw the line between 3. and 4. The API change in 3. doesn't fundamentally change how Nushell works.

## Type-safe eval

``-
eval spam.nu [
foo [x: string]: int -> list<string>
bar [--flag]
]

spam foo
spam bar

eval eggs.nu [
baz [x: string]: stream<int> -> anyof<stream<int>, stream<string>>
] *

baz

Use cases:
* plugins (so we don't need to talk to externals during parsing)
* maybe test runner