[GeophyAI]

Every example was run as documented after the merge into one package. The renames themselves broke nothing; these commits fix what did not run or did not match its docs.

• Viking: the conventional SIREN/hash, RTM and six sbatch examples depended on files that no longer exist. They are replaced by one frequency-selection FWI example (all 1001 shots on their own comb bins, six bands to 16 Hz) with a pipeline config, an sbatch template and a rewritten walkthrough. The starting model is downloaded from the Hugging Face copy of the data set.
• Tasks: 03/07 no longer run out of memory on a 32 GB GPU at the stage obs bandpass; $VAR is expanded in every YAML string and an unset variable is an error (task YAMLs and --config files); the init templates run as written; LSRTM template lr; RTM checks grid vs model shape and fails when every batch fails.
• Synthetic example docs: a CPU fallback that loads and fits in memory; headers and pages corrected to measured results.
• sweepx.opt: line-search optimizers move m/s models with their default step; strong Wolfe rejects non-finite trials and projects onto bounds; every docs/opt snippet runs. DepthWeighting now defaults to depth_axis=0 (depth is axis 0 in sweep models).

Tested: full tests/ on a V100 (1851 passed, 2 skipped) and on CPU; mkdocs build gives no new warnings; the Viking walkthrough end to end on a V100 (52 min inversion, misfit 0.484 -> 0.188).

🤖 Generated with Claude Code