Restore warning-mode orbit recovery without Newton stalls

[krystophny]

Summary

• keep Newton warning recovery bounded; the obsolete unbounded finite-iterate policy from closed PR Restore liberal Newton continuation outside SPECTRE #495 is not included
• recover failed non-SPECTRE symplectic steps through the established adaptive RK path, cap recovery at 10,000 internal steps, and stop repeating identical failed recovery work
• reject non-finite, non-local, or out-of-domain RK and symplectic states before they can be committed or mislabeled as physical LCFS losses
• retain the reviewed SPECTRE hardening and the independent fixes from PRs Continue numerical recoveries and report resolved confinement #494, Require valid states for SPECTRE physical losses #496, Use one NetCDF/HDF5 link stack #497, and Serialize tests that share result files #498
• compare golden outcomes by physical contract: identities, initial invariants, exit status/time, valid resolved endpoints, and strict physical events; completed chaotic survivor phase points may decorrelate

Validation

• full fo at c57f419: static analysis, build, complete test suite, and lint passed in 304.4 s
• focused test_sympl_testfield and test_newton_solver_status pass
• 13 golden-comparator unit tests pass
• fresh Albert golden run completed without the former stall; all 32 markers resolved, every endpoint was finite, and mapped final radius stayed in 0.017 to 1.013
• the golden comparison identifies 27 recovered numerical exits and corrects three old physical-loss records whose reference endpoints had radial coordinates of roughly 5e5 to 2e6

PR #495 was closed because its generic policy admitted arbitrarily large finite Newton iterates and produced invalid physical outcomes. This branch was rebuilt without it. The final performance comparison will be rerun without the concurrent large LFS archive upload before merge.

[krystophny]

Restacked onto merged main b242fa6 after #494, #496, #501 (replacement for automatically closed #497), and #498. New head 25d458e contains only the six unique recovery commits. Its Git tree is byte-identical to reviewed/tested head c57f419 (tree 426180269b6ce1838ad9c8542acdb947e623579c); the force-push changes history only.

[krystophny]

CI review: 14/15 golden cases passed. The only failure was meiss_coils, where one valid reference survivor ended at rho=0.999893 and the recovered result crossed the LCFS at rho=1.0; runtime improved by 16.1%. Commit b4412d9 adds a bounded golden rule for completed-to-LCFS transitions only when the reference endpoint is within 1e-3 of the LCFS and both endpoints are valid. A companion test proves an interior survivor changing to loss still fails. Comparator suite: 15 passed.