[FFroehlich]

Bumps the version to 0.8.2. After merging, publishing a GitHub release with tag 0.8.2 triggers the PyPI deploy.

Release notes (to be used for the GitHub release):

---

Bug-fix release. It fixes the open items from the adversarial review in #75 (#77–#83).

Behaviour changes

• A trial point with a finite function value but a non-finite gradient or Hessian is now rejected and the trust region shrinks, as for a non-finite function value. Before, minimize() raised RuntimeError. Each such step logs a warning. Non-finite values at x0 still raise (Trial point with finite fval but non-finite gradient raises instead of rejecting the step #79, Reject trial points with non-finite gradient or Hessian instead of raising #85)
• subspace_solver='scg': the Steihaug CG tolerance is now relative to the gradient at the first iteration, min(0.5, sqrt(|g| / |g0|)) * |g|, so it no longer depends on the scale of f. Fewer runs stop with FTOL short of the optimum on ill-conditioned problems. Iteration counts and results change for existing scg users (scg: Steihaug forcing tolerance gives steepest-descent steps and false FTOL on ill-conditioned problems #77, scg: make the Steihaug CG forcing term invariant to the scale of f #89)
• The trust region now also grows after a stepped-back step with a good ratio if the step before step-back reached the trust-region radius. reflect_single no longer crawls to maxiter when several variables sit on bounds (reflect_single crawls to MAXITER when several variables sit on bounds #78, Grow the trust region for stepped-back steps whose proposal hit the radius #90)
• Loggers are now named fides.<n> instead of fides_<n>, so they are children of the fides logger. Default output is unchanged. To avoid duplicate lines after logging.basicConfig(), set logging.getLogger('fides').propagate = False (One logger per Optimizer instance accumulates in the logging manager #81, Pool per-optimizer loggers instead of leaking one per instance #86)

Performance

• stepback_strategy='refine' solves its subproblem with a new active-set solver instead of scipy trust-constr. It now takes about as long as reflect (was 10–100× slower), and its steps are as good or better (refine step-back is 10–100× slower than the other strategies #80, Solve the refine step-back subproblem with an active-set method #88)
• stepback_reflect stops reflecting once a reflected step bounces back at the same bound with a negligible segment, which saves subproblem solves on problems with many active bounds (stepback_reflect keeps reflecting down to alpha ~ 1e-16 #83, Stop reflecting once reflected steps bounce with negligible segments #87)

Bug fixes

• options={'history_file': None}, the documented default, raised TypeError. history_file now accepts None, str and any os.PathLike (options={'history_file': None}, the documented default, raises TypeError #82, Accept None and os.PathLike for the history_file option #84)
• Creating one Optimizer per start (as pyPESTO does) leaked one logger and handler per instance. Loggers are now pooled and reused once their Optimizer is garbage collected (One logger per Optimizer instance accumulates in the logging manager #81, Pool per-optimizer loggers instead of leaking one per instance #86)
• The tracked minimum (x_min, grad_min) can no longer be a point with a non-finite gradient (Trial point with finite fval but non-finite gradient raises instead of rejecting the step #79, Reject trial points with non-finite gradient or Hessian instead of raising #85)

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 97.13%. Comparing base (a3d4722) to head (dbb1c71).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main      #91   +/-   ##
=======================================
  Coverage   97.13%   97.13%           
=======================================
  Files          10       10           
  Lines        1256     1256           
=======================================
  Hits         1220     1220           
  Misses         36       36           

☔ 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.