Harden core pipeline: input-robustness and correctness fixes (ranking, file IO, isomer/SPE)

[isayev]

Summary

A systematic per-module hardening pass over the core conformer-generation pipeline. The recurring root cause addressed: newer functions had input guards (None/blank/duplicate) that their older siblings lacked, so a single bad record — a None molecule, a blank line, a missing property, or a salt/metal atom — could crash an entire run or silently drop data. 16 confirmed bugs fixed across 6 modules, each with a regression test.

Ranking & filtering

• top_k with k=1 (the default --k=1) now validates connectivity instead of emitting a structurally-broken lowest-energy conformer.
• run() no longer aborts the whole ranking when one record lacks the Converged property.
• check_connectivity is salt/metal-safe (previously KeyError on elements outside its radii table).
• Legacy filter_unique treats an RMSD-comparison failure as "distinct" (keep) rather than "duplicate" (drop), matching the optimized filter.

File IO (file_ops, chunk_manager)

• None-molecule guards in decode_ids, reorder_sdf, find_smiles_not_in_sdf, count_sdf.
• reorder_sdf is now crash-safe (temp file + atomic replace) and data-preserving — this also closes a real smiles2mols bug where duplicate input SMILES silently dropped a conformer.
• SDF2chunks keeps a terminator-less trailing molecule; smiles2smi raises an actionable error on invalid SMILES.
• Blank-line, ragged-.smi, and degenerate-chunk_size guards.

Isomer enumeration & single-point energy (isomer_engine, SPE)

• One invalid SMILES no longer aborts the whole batch (guards at four sites).
• Stereoisomer enumeration no longer silently truncates at 1024 isomers (explicit cap + truncation warning).
• Clash filter: closed the boundary dead-band and added an MMFF→UFF fallback (rescues B/Se-containing molecules).
• Conformer-count budget unified across the SMILES and SDF paths on the with-H representation (restores richer hydroxyl/amine sampling) and floored at 1.
• SPE filters None/conformerless records and fixes an output index-misalignment; actionable error for unsupported ANI2xt elements.

Test plan

• Full fast suite: 595 passed, 45 deselected (was 563 on main; +32 regression tests, all fail-before/pass-after).
• ruff check src/ --select F401,F841,UP007 clean.
• Each fix independently reviewed (chemistry/correctness) before commit.

Notes

• Branches from main (post-3.5.0). No behavior change on the normal pipeline path beyond the documented correctness fixes; reorder_sdf's normal (all-matched) ordering is byte-identical.
• Separate, pre-existing issue (not in this PR): the slow-gated test_calc_spe_userNNP2 fails because torch.jit.script cannot compile the aimnet AEV module when wrapping a torchani-based custom NNP — worth a dedicated follow-up.
• Remaining audited-but-unfixed items queued for follow-up: stereochemistry public-API SMILES surgery on explicit stereo tags, ANI2xt fp64 energy-shift precision, and pad_from_mols host→device transfer batching.