[Sophia Wen (hfwen0502)]

A pass over every README, from the top level down, and the docs front page, so that each reads as how to use the package rather than as a record of upstream's internals.

Why use the bindings. The README gains a short section on what the bindings give over upstream's C++ apps: one install for every feature, the backend and the unrestricted and complex builds chosen per call rather than at compile time, Python objects in and out, and an error naming the right device when one does not support a feature. The docs front page gains the same, with the README's feature list and its feature-by-backend table, and the sbd API page's intro now covers CAOP and device=.

Commands use the hardware. Examples launch on several ranks, or one rank per GPU, wherever the step parallelizes: TPB on a rank grid or across GPUs, CAOP on GPUs, GDB sharded. Commands stay single-process only where the step is serial.

Trimmed. Long explanations of upstream internals are cut to what a user acts on, and repeated warnings appear once. The GDB README's command blocks list one example per feature and point to the section that explains each option.

The analysis README is organized by use case: checking a molecular state, checking a CAOP state, and estimating the exact energy from a series with extrapolate.py, which it previously mentioned only in passing. It opens with which upstream evaluator each kind of state uses, gdb-stat-evaluator or caop-stat-evaluator. The TPB-versus-GDB growth comparison is dropped: it compared solvers rather than showing the analysis.

Also:

• Why the SQD loop is TPB only, citing Support general non-Cartesian product subspaces qiskit-sqd#389.
• How a UHF FCIDUMP selects the unrestricted build, that UHF runs on gpu too, and that the variance analysis takes restricted and UHF FCIDUMPs alike.
• The examples are tested against the pinned upstream commit only; with another, the functionality may be broken.
• The workflow page no longer says PT2 recovers most of the missing energy over a too-small CAOP basis; it recovers only part.
• The GDB README's --initial_det example runs as written.
• SBD_BUILD_UHF is removed from INSTALL.md, the release notes and setup.py, the one change outside the docs: a three-line warning for a variable no release had, so the next release neither introduces nor removes it.

Checked:

• The commands added or changed run on an 8× H100 server, apart from two long heat-bath runs stopped before they finished.
• The variance analysis runs on saved UHF and CAOP states.
• Every relative link and section anchor resolves.
• The docs build with -W.

🤖 Generated with Claude Code