feat: implement __sub__ and __rsub__ for Loss class

[hemv-857]

What

Implements __sub__ and __rsub__ on the Loss class so users can write:

combined = SomeLoss() - OtherLoss()
constant_sub = 5 - SomeLoss()

Previously the only way to achieve this was SomeLoss() + (-1) * OtherLoss().

Closes #68

Changes

• Added __sub__ method to Loss class (follows __add__ pattern exactly)
• Added __rsub__ method for constant - loss syntax
• Added sub to parametrized arithmetic tests
• Added sub to repr arithmetic tests
• Added type error tests for subtraction with invalid types

Testing

All existing tests pass. New tests cover:

• loss_a - loss_b (Loss - Loss)
• loss - constant (Loss - int/float)
• constant - loss (int/float - Loss via __rsub__)
• TypeError on invalid operand types
• Repr round-trip for subtraction

[Copilot]

Copilot review overview

🟡 Changes recommended

Unresolved moderate issues affect unary negation and nested-subtraction repr correctness; reverse-subtraction coverage and documentation also need updates.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 2 · 1

Open (3)

• Unary negation lacks neg support · New
• Right-hand composite subtraction loses grouping · New
• Parametrization omits valid reverse-subtraction coverage · New

What changed in this PR

Adds subtraction support to Loss, including reflected constant subtraction, with expanded arithmetic and representation tests.

Changes:

• Implements Loss.__sub__ and Loss.__rsub__.
• Adds subtraction arithmetic, repr, and invalid-type test coverage.

File	Review summary
tests/​test_losses.py	Adds subtraction coverage, but lacks valid reverse-subtraction evaluation and representation tests.
deepdow/​losses.py	Adds subtraction operators, but unary negation is unsupported, nested subtraction repr loses grouping, and the docstring does not match supported operand types.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[hemv-857]

Thanks for the review! All findings addressed:

• Unary negation: Added __neg__ so -SomeLoss() now works. repr uses -() wrapping.
• Repr grouping: Right operand in __sub__ is now parenthesized: a - (b) instead of a - b. This preserves eval(repr(...)) round-trip for nested subtraction like a - (b - c).
• rsub test coverage: Added test_rsub that exercises constant - loss with evaluation and repr round-trip checks.
• Also added test_sub_repr_roundtrip specifically for nested subtraction grouping.

[Copilot]

Copilot review overview

🟡 Changes recommended

A test has undefined class names, and reflected subtraction mishandles Loss subclasses.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 1

Open (1)

• Use defined loss class names in subtraction repr test · New

Resolved since last review (3)

• Right-hand composite subtraction loses grouping
• Unary negation lacks neg support
• Parametrization omits valid reverse-subtraction coverage

Previously missed (1)

In code that hasn't changed since last review

Handle reflected subtraction for Loss subclasses

deepdow/​losses.py:369

__rsub__ rejects Loss operands, but Python can dispatch the reflected method for left_loss - right_loss when right_loss is a subclass of left_loss. In that case a valid Loss–Loss subtraction reaches this branch and raises TypeError; handle Loss by delegating to other.__sub__(self) before the numeric case (and add a regression test with a derived loss class).

[hemv-857]

Thanks for the review! Both findings addressed:

• Undefined class names: Fixed Variance → StandardDeviation and MaxDrawdown → MaximumDrawdown in test_sub_repr_roundtrip. Test now uses the actual imported class names.
• Loss subclass dispatch in rsub: Added isinstance(other, Loss) check to __rsub__. When base_loss - subclass_loss and the subclass doesn't define __sub__, Python dispatches to subclass.__rsub__(base_loss). Now handles this case correctly. Added test_rsub_with_loss_subclass regression test.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[hemv-857]

Hi! Just checking in on this PR — happy to make any changes if needed. Thanks for your time!