Action the review of the agent instruction files - #7347

Separate "never import *" from the __all__ rule, and drop the ban on
conditional imports in a package __init__.py. The ban sat under
Greppability, where it does not belong: a try-wrapped import of a verbatim
name greps perfectly well. It also contradicted the tree, where iris_grib
and iris.fileformats.zarr both defer an optional dependency's import.

Record that indexing does not shrink a Dask graph. A slice still references
every chunk of its source, and is culled only at compute.

State the purposes of a docstring, a comment and a module docstring at the
head of the section that governs all three.

Replace the guard-clause advice with a single-exit rule. Nearly every
function in lib/iris that contains a return has exactly one, so the old
wording argued against the surrounding code.

Delete the prose describing the sphinx-needs item a public module docstring
carries. The check_docs_page_metadata hook enforces it, and names the
expected type, the line limit and the missing tag when it fails.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

A module docstring is the only prose read on every visit to a module and
reviewed in the diff, which makes it the right home for subsystem
explanation. Sphinx publishes it, so prose written for a maintainer arrives
in front of a user. This directive drops its content instead, leaving it in
the file and off the page.

Built on docutils and sphinx alone, so it adds no dependency. The content is
never parsed, which makes broken markup and unbalanced doctest prompts safe
to write and keeps the doctest builder away from them. Misspelling the name
stops the build, because make html runs with -W, and the content does not
render on the way out. A plain reST comment fails the other way: dedent one
line and the prose is silently published.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Seventeen module docstrings now split by reader. What a caller needs stays on
the page; design rationale, traps, invariants and coupling move into a
maintainer-notes block that Sphinx drops. Two of the seventeen are the saving
modules added by SciTools#7346, which never carried a marker.

iris.io.format_picker loses its marker without gaining a block. Its docstring
is a usage tutorial throughout, so it was marked on length rather than on
content, and nothing in it is written for a maintainer.

Two corrections made in passing, both to text that was being moved anyway.
iris.fileformats.zarr._dataset announced three things about Zarr and then
listed four. iris.fileformats.cf pointed at "the plan, section 8.4"; section 8
of the design document is Risks and has no subsections, so that resolved to
nothing and is dropped.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Session-Id: 76a82181-7fe1-4063-9fd7-deb10fb0d658

The non-atomic clobber behind the "w-" default is a data-loss warning, not a
rationale: a user choosing mode="w" needs to know an interrupted Zarr clobber
destroys the old store without completing the new.

iris.fileformats.netcdf.loader had one thought split across the boundary.
What is left in the module and what moved to the CF loader is one statement,
so it goes back whole, into the block.

Naming only the netCDF saver implied it was the only one; there are two
public _CFSaver subclasses.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Session-Id: ee18e927-ba60-4705-aee1-029036b34706