Remove intercept for a bare "+ -1" / "+ 0", matching "- 1" (fixes #158)

[deepakganesh78]

This fixes #158.

Root cause

_eval_binary_plus only propagated an added intercept from its right operand (via the intercept flag); it ignored the right operand's intercept_removed flag entirely, keeping only the left operand's intercept state.

So a + b - 1 removes the intercept (the removal lives on the left operand), but a + b + -1 did not — even though the docs imply that appending + -1 is equivalent to appending - 1. The maintainer confirmed in the issue that this is a bug and that the successor, Formulaic, treats a + b + -1 as a + b.

Fix

When the right operand of a binary + is a bare intercept literal (a chain of unary +/- ending directly in 0/1, as in -1), honor its intercept removal — mirroring the existing direct + 0 (ZERO) handling just above it.

A small _is_bare_intercept() helper inspects the parse node so that removal is not propagated out of a compound/parenthesized expression such as 1 + (a + 0), which keeps the intercept by design (covered by the existing ~ 1 + (a + 0) test, which still passes).

Behavior after fix

a + b - 1      -> ['a', 'b']
a + b + -1     -> ['a', 'b']   # was ['Intercept', 'a', 'b']
a + b + 0      -> ['a', 'b']   # was ['Intercept', 'a', 'b']
a + -1 + 1     -> ['Intercept', 'a']   # re-adding still works
0 + a          -> ['a']        # unchanged
1 + (a + 0)    -> ['Intercept', 'a']   # unchanged, by design

Tests

Added regression cases to _eval_tests and a changelog entry. Full suite is green:

147 passed, 1 skipped

(scipy installed so the spline tests run.)

[bashtage]

It would be appropriate to warn and let the user know since they can easily fix it.

[bashtage]

In my opinion, this is a "won't fix" type of bug given where the project is. I think a warning would be OK, though. To fix this, one would need to go through a reasonably long deprecation cycle since this is a change to existing code, even if the original implementation is not great. Something that explains to the user what happens here and then tells the user how they can fix it (by slightly altering their formula) is better than a silent behaviour change.

statsmodels will be out with formulaic support next month, so this is moving to lower importance.