[SteveDiamond]

Three bugs found while auditing the codebase ahead of the patch release. Each is covered by a new regression test.

🔴 1. prob.value had the wrong sign for maximization problems

The dccp() return value was correctly de-negated for Maximize, but the value written to prob.value was the raw (negated) subproblem value. So a solved maximization problem returned the right number while prob.value held its negation.

Maximize(norm(x)) s.t. 0<=x<=1   # optimum sqrt(2) ≈ 1.414
  before:  return 1.414   prob.value -1.414  ❌
  after:   return 1.414   prob.value  1.414  ✅

Fixed in both the single-init (_solve) and multi-init (solve_multi_init) paths by setting prob.value from the original objective after writing the variables back, so the sign matches the problem sense.

🔴 2. Infinite loop in constraint convexification damping

The constraint-convexification damping loop was unbounded (while c_conv is None:), unlike the objective path which is bounded by max_iter_damp. When damping could not recover a valid linearization point (e.g. a variable lands outside a function's domain such as sqrt/log and _apply_damping is a no-op for it), the solve hung forever. Now bounded by max_iter_damp and raises NonDCCPError, mirroring the objective path. Extracted into _convexify_constr_with_damping.

🟡 3. seed ignored for multi-restart runs

_solve_multi_sequential / _solve_multi_parallel called initialize(..., random=True) without passing the configured seed, so seed= did not make multi-restart (k_ccp > 1) runs reproducible. Restarts now derive a distinct, reproducible per-restart seed via _restart_seed.

Testing

• uv run pytest: 85 passed, 1 skipped (was 63; added 4 regression tests)
• uv run ruff check: clean
• uv run pyright: 0 errors

🤖 Generated with Claude Code

[SteveDiamond]

@d-002 these are other bugs that I will fix in the v1.1.1 patch release. If you have anything else you would like fixed besides the bug you already identified, post here.

[d-002]

Hi,
Thanks for the mention, I did notice another warning (RuntimeWarning: Reading from a sparse CVXPY expression via ".value" is discouraged. Use ".value_sparse" instead) when using the latest commit as a package.
I haven't had time to investigate, but maybe this could be worth fixing as well?

Edit: after a discussion with Gemini, it appears this warning could be suppressed, as it means CVXPY converted a sparse array into a dense one in a specific case where it could lead to loss of performance.
However, the DCCP codebase mainly uses dense numpy arrays so the warning is not relevant.

[SteveDiamond]

Thanks! The warning turned out to be only fixable on the cvxpy side, and it's also spurious (just checking if something is None). But I'll create a cvxpy issue about it. For the moment I suppressed the warnings.