[mergify]

A scan of everything the docs site renders (the landing page, the three API pages and the 22 docstrings they pull in) turned up statements that are out of date or misleading. Each was checked against the code. This fixes them. Docstrings and one table row only; no code changes.

Wrong

• gpu-omp was missing from the device lists. tpb_diag, tpb_diag_from_files and gdb_diag documented device as ('cpu', 'gpu', or None), and get_backend as 'cpu', 'gpu', 'auto', or None. All four pass device to get_backend, which accepts 'gpu-omp' and its aliases. gdb_diag now also says that 'gpu-omp' runs GDB on the host, since SBD has no OpenMP-offload GDB code.
• get_backend(None) does not mean "the default". It resolves to _default_device or 'auto', so before anything has initialized it is 'auto', which prefers Thrust. Meanwhile the first tpb_diag or gdb_diag call initializes with 'cpu'. The docstring now says this and recommends calling init() to avoid depending on call order.
• The bit_length row limited the odd-value hazard to GDB's heatbath expansion. Upstream's own diag also gives a wrong plain-diagonalization energy at --bit_length 31 on its bundled Fe4S4 data (GDB app: odd --bit_length gives a different energy or a crash (Fe4S4). Is an even value expected? r-ccs-cms/sbd#104), so the row now covers all GDB use with multi-word determinants.

Misleading or incomplete

• tpb_diag_from_files takes one determinant file and uses it for both alpha and beta (bdet = adet upstream); the docstring now says so, and points to tpb_diag for different lists.
• init now lists gdb_diag among the calls that take a per-call device.
• get_backend said passing device is how you switch between CPU and GPU in one process. That holds except for the CPU backend together with gpu-omp, the combination the 1.7 release notes warn about; it now says so.
• assemble_rdms cited a private repository for SBD's RDM layout. It now cites upstream SBD's docs/user-guide.md, which documents exactly that layout, and no longer refers to code "below".
• from_strings, get_backend and DeviceConfig.gpu_omp linked to from_string, makestring and has_backend_conflict, none of which is on the API page, so they rendered as plain text. They are now literals, which adds no API surface.
• DeviceConfig's usage example called the undocumented apply() on a TPB_SBD object named sbd_config, a name that means a dict elsewhere. It now passes the config to solve_sci_batch(device_config=...).

Minor

• TPB_SBD, GDB_SBD, LoadFCIDump and LoadAlphaDets get real docstrings. LoadAlphaDets documents its arguments, and that it neither sorts nor deduplicates.
• DeviceConfig.auto drops development history and a reference to a private function.
• The from_strings examples use bit_length 62 instead of 63; the new outputs were checked by running them and round-tripping through makestring.

sphinx-build -W -T --keep-going, the command tox -e docs runs, succeeds with no warnings, and every new passage appears on the rendered pages. Labelled for backport, because the site deploys from stable/1.7.

🤖 Generated with Claude Code

---

This is an automatic backport of pull request #60 done by [Mergify](https://mergify.com).