[FBumann]

Prompt: What about the last bug in lpspec? … Full fix, both shapes.

Note

The following content was generated by AI.

A constant piece added beside a variable term is now asked for its data on both
lanes. x + hi <= 100 with hi short of a coordinate read x <= 100 on the
eager lane — the missing hi filled with the zero that is a bound — while the
relational lane refused; the two lanes disagreed, and the eager one built the
wrong constraint in silence.

Both lanes' constant-side check skipped any side that carried a variable, so a
piece beside a term went unasked. They now ask of every constant piece — a
parameter in a variable-free additive position — found by threading a
coefficient flag through the walk: a parameter a variable stands with in a
product is a coefficient, whose missing row is the zero the absence rules allow,
and is left alone.

Closes #1521.

The shapes, and both lanes' verdict

constraint, hi/c short one coordinate	before	after
x + hi <= 100	eager builds x <= 100, relational refuses	both refuse, one sentence
sum(x) + sum(hi) <= 100	both build (the gap is summed away)	both refuse, one sentence
(x + hi) * 2 <= 100	eager builds, relational refuses	both refuse
w * x <= 100, w short	both build (w=0 is a legal coefficient)	unchanged — still both build

The aggregated shape (sum(x) + sum(hi)) is the one #1465 left, missed on both
lanes; the eager check asked of a variable-free side and the relational one
could not see the hole through the sum. Both now ask the parameter directly.

What was verified

• Three tests in tests/test_absence.py: both lanes refuse w * x + c <= 100
  and its summed form with one identical sentence (both_lanes_refuse), and a
  sparse coefficient beside a covered piece still builds and agrees
  (differential(..., lp=True)). The two refusal tests fail on the tree before
  the fix (the eager lane built in silence); confirmed.
• Full suite: 3684 passed, 329 skipped; the 10 failures are all gurobi /
  xpress cases, which need commercial solvers absent from this environment —
  94 pass with HiGHS across the two affected files. ruff check /
  ruff format --check clean; the changed functions add no pyrefly error.
• Not run: pixi run check — pixi is unavailable here, so gates ran through
  a local venv.

Mutation table

Each guard deleted in turn, the affected tests run (tools.mutate needs a
committed tree it did not have mid-work, so taken by hand):

guard removed	caught by
the if not coefficient: yield-suppression (both walks)	test_a_sparse_coefficient_is_still_a_zero_coefficient, …_beside_a_constant_piece_is_still_a_zero, …_on_the_bound_side_still_pins_the_variable fail
the coefficient or carries_variable(…) propagation through a product	the same three coefficient tests fail
the widening (restore the whole-side carries_variable skip)	test_a_constant_piece_beside_a_term_is_refused_on_both_lanes, …_through_a_reduction fail

On the stack

Meant as a PR stacked on #1536's branch, but #1615
merged to main first, so this targets main directly — the base already
carries #1536, and the diff is this fix alone.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JyxyehhHvXsBqf6NxeQgU1

---

Generated by Claude Code

[codspeed]

Merging this PR will not alter performance

✅ 24 untouched benchmarks
⏩ 58 skipped benchmarks¹

---

_{Comparing claude/constant-piece-beside-a-term (6f64147) with main (089f816)²}

Footnotes

1. 58 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

2. No successful run was found on main (59391f3) during the generation of this report, so 089f816 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩