Refactor Instrumentation to Remove "Magical" Builtins Injection

[apiad]

Description:

The current CLI implementation relies on injecting the _ferret_profiler instance directly into the builtins module to make it available to the instrumented AST. This is a "magical" and fragile approach that poses significant stability risks:

1. Namespace Pollution: It modifies the global python environment, which can conflict with user code or other libraries that might use a similar variable name.
2. Shadowing Risks: If a user script accidentally defines a variable named _ferret_profiler, the instrumentation will crash or behave unpredictably.
3. Refactoring Difficulty: It creates a hidden, implicit dependency that makes the codebase harder to test and maintain.

Proposed Solution:

Replace the builtins injection with a robust Singleton Accessor Pattern. The FerretInstrumentor should inject explicit import statements into the user's AST that retrieve the active profiler from a safe, module-level accessor.

Tasks:

• Core Refactor (ferret/core.py):
  • Implement a module-level global variable _global_profiler.
  • Add thread-safe accessor functions: set_global_profiler(profiler) and get_global_profiler().
• Instrumentor Update (ferret/cli.py):
  • Modify FerretInstrumentor to stop assuming _ferret_profiler exists in builtins.
  • Update the _inject_decorator logic to insert a localized import at the top of the instrumented function or module (e.g., from ferret.core import get_global_profiler).
  • Change the decorator call to use this imported accessor: @get_global_profiler().measure(...).
• Expanded AST Support:
  • Extend FerretInstrumentor to visit ast.Lambda nodes and wrap them in a profiling context, as these are currently ignored.

Acceptance Criteria:

• Running ferret run script.py works without builtins._ferret_profiler being present in the locals() or globals() of the user script.
• User scripts that define their own _ferret_profiler variable do not cause the profiler to crash.