test: real protocol assertions and an exercised enable_runtime_checks

[thomaspinder]

Summary

Two test-infrastructure issues, one PR — and the new test immediately earned its keep by surfacing two real defects.

• test: tautological runtime_checkable assertions in test_protocols.py #174: five always-true assertions in tests/test_protocols.py (assert hasattr(P, "__protocol_attrs__") or runtime_checkable — the second operand is a truthy function object) replaced with real ones: typing_extensions.is_protocol + exact get_protocol_members sets per protocol (py3.11-verified — stdlib __protocol_attrs__ is 3.12+), plus conforming/non-conforming stub isinstance checks that genuinely pin runtime-checkability.
• test: enable_runtime_checks is never exercised beyond importability #181: enable_runtime_checks() was only ever asserted callable. A new subprocess test enables it in a fresh interpreter, drives a full synthetic pipeline (VARData → VAR → FittedVAR → sigma/forecast → IdentifiedVAR → IRF), and asserts a BeartypeCallHintViolation actually fires for a mistyped call. It failed immediately, exposing two pre-existing defects that made the feature unusable (exactly as the feat(diagnostics): VAR-aware convergence and stability report #176 review predicted):
  1. Beartype resolves stringified annotations (-> "ForecastResult", "xr.Dataset") against the defining module at call time, and every such name was TYPE_CHECKING-only → BeartypeCallHintForwardRefException on first use. Fixed with _bind_deferred_imports() — AST-parses each wrapped module's if TYPE_CHECKING: block and binds the imports for real (no exec), including dotted aliases.
  2. The wrap loop decorated every class in dir(mod) — including typing.Protocol and classes owned by other modules (Constant reached through spec), mutating shared state. Fixed with a __module__ == mod.__name__ ownership guard.

Test teeth verified: with the src fix stashed, the subprocess test fails with the original forward-ref error.

Closes #174
Closes #181

Gates

ruff / ty / fast suite green (532 passed, 29 deselected); the rewritten protocol file passes standalone under Python 3.11.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Egjd7ToFeb9TQqFnfRQZxV

[codecov-commenter]

Codecov Report

❌ Patch coverage is 4.34783% with 22 lines in your changes missing coverage. Please review.
✅ Project coverage is 93.3%. Comparing base (fa3375c) to head (4175dc5).
⚠️ Report is 2 commits behind head on main.

Files with missing lines	Patch %	Lines
src/impulso/__init__.py	4.3%	22 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##            main    #209     +/-   ##
=======================================
- Coverage   94.1%   93.3%   -0.8%     
=======================================
  Files         41      41             
  Lines       2479    2503     +24     
  Branches     290     298      +8     
=======================================
+ Hits        2333    2336      +3     
- Misses        98     119     +21     
  Partials      48      48             

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.