[HazenBabcock]

Makes the repository easier to work in, whether the reader is a person arriving
for the first time or an automated tool. Five commits, one concern each. No
analysis code changes; 280 tests green and the documentation builds without
warnings.

Line endings

Twenty-eight files were committed with CRLF while the other 579 text files used
LF, spread across ten directories with no rule to infer - admm_lasso.c is CRLF
and admm_lasso.py is not. Nothing depended on it, but any tool that rewrites a
file whole converts the endings as a side effect, so a two line change arrives as
a diff touching all 765 lines. That happened repeatedly during the recent audit
work.

The conversion is CRLF to LF and nothing else:

• stripping every CR from each file before and after gives identical bytes
• git diff --ignore-cr-at-eol is empty across the whole change
• insertions and deletions are equal at 7140 each, as they must be

.gitattributes keeps it from returning. text=auto lets git detect binaries
itself, with an explicit list underneath as a backstop - the checked-in Windows
DLLs matter most, since rewriting those would break the wheel the Windows job
builds. A fresh clone now comes out with zero CRLF files.

AGENTS.md

There was nothing describing how to work on this project. The new file covers
build, tests, running an analysis, diagnostics, layout, conventions and traps,
and every count and path in it was checked against the tree.

It records the things that are only learnable by getting them wrong once:

• re-run scons after editing C, because the Python side loads whatever is
  already in c_libraries/ and a stale library gives a plausible wrong answer
  rather than an error
• trust pytest's exit status - several C libraries flush stdout at interpreter
  exit and push the N passed line off the end of the output
• pytest does not run the __main__ blocks, which is how a renamed test stayed
  green across six pull requests
• decon_storm/ has C sources that SConstruct never builds and nothing
  imports, so a definition found there is not the one that runs

The conventions section lists only what was measured: all 242 modules that
import numpy write import numpy, and as np appears nowhere; class methods
are camelCase 704 times against 6 in snake_case. Keyword argument spacing is
explicitly called out as not a convention - it is close to evenly split and
108 of 254 files use both forms, so tidying it either way would be inventing a
rule.

PSF measurement input files

This is a real documentation defect rather than a polish item. analysis.rst
said to create "a text file containing the z value of each frame in the z stack
movie". measure_psf_beads.py passes that file to makeZIndexArray(), which
asserts a shape of (N,2): a valid flag and the z offset in microns, with
flag-0 frames dropped. Following the instructions as written ends at

AssertionError: Z offsets must have shape (N,2).

with nothing to indicate what the second column is. offset_to_z.py writes the
right format but was mentioned only in the Multiplane section a hundred lines
below, and it reads a storm-control .off file so it is not a general answer.

The bead location file had a smaller version of the same problem - described
only as "a text file containing the locations of a few (isolated) beads", with
the format living in a source comment. Both now have a stated format and an
example, verified against the files the spliner diagnostic actually generates.
Multiplane's step 5 takes the same two columns and now refers to the Spliner
description instead of restating it loosely.

Also fixes beads_zoffset.text, spelled .txt two lines below and everywhere
else.

Running an analysis

AGENTS.md covered changing this code and said nothing about driving it, though
"analyze this movie with 3D-DAOSTORM in 2d fixed mode" is an ordinary request.
Running one works as documented; what is missing is the two things that go wrong
quietly.

Parameter files are always copied from an existing one, and four of the values
in them describe the microscope rather than the analysis - pixel_size,
camera_gain, camera_offset and the initial sigma. Nothing compares them
against the movie, so inheriting them from a sample gives a clean run, a full
set of localizations and the wrong units.

The other is that parameter problems are reported at the top of the output. An
unrecognized parameter already prints a warning, correctly, as the first line -
followed by several hundred lines of per-frame progress, so anything reading the
tail concludes the run was clean.

Diagnostics README

e644347e said "the rest take no arguments" when describing --psf-model.
fista_decon, spliner, spliner_2d and multiplane all accept
--no-splines. Only the missing --psf-model actually stops a script, so the
note now distinguishes required from optional.