Add an autouse fixture that snapshots and restores Logger._reaction
so the WARN mode set by _warn() cannot leak into later raise-expecting
tests (e.g. core/test_guard, core/test_parameters).

Add and extend supplementary coverage tests for analysis (analysis,
fitting, minimizers/base), display (plotting, progress), report
(data_context, html_renderer), background estimate, and the package
__main__ entry point. Raises overall unit coverage to ~84.5%.

Supplementary unit tests lift tex_renderer 67->100%, iucr_writer
84->97%, cif serialize 83->96%, and fit tracking 78->98%. Overall unit
coverage rises to ~86.3%.

Supplementary unit tests lift project/display 73->100%, utils/logging
76->99%, utils/utils 86->99%, and experiment item/base 76->99%. Overall
unit coverage rises to ~87.8%.

Supplementary unit tests lift display/plotting 82->93%, analysis 82->94%,
plotters/plotly 78->97%, and project/project 80->100%. Overall unit
coverage crosses 90% (90.0%).

Supplementary unit tests lift analysis/sequential 53->99%,
utils/environment 69->95%, io/results_sidecar 81->99%, and experiment
collection 47->100%. Overall unit coverage rises to ~91.4%.

Document previously-undocumented methods, properties, setters, and
__init__s across analysis (categories, minimizers, fit_helpers,
calculators), datablocks/experiment (peak, instrument), and display.
Raises docstring coverage from 86.2% to 94.4% (interrogate).

_plot_single_crystal_posterior_predictive_summary called
PlotlyPlotter._get_diagonal_shape() with no arguments, but the method
requires (minimum, maximum); the reflection-check plot raised TypeError
whenever reached. Compute the shared axis range via
single_crystal_axis_range (matching build_single_crystal_figure) and
pass it through. Adds a test covering the now-working build path.

test_relative_destination_does_not_depend_on_current_project relied
on a sibling test leaving EASYDIFFRACTION_ARTIFACT_ROOT cleared; it
failed in isolation because the configured artifact root rooted the
relative destination under tmp/tutorials instead of the cwd. Clear the
env var in the test so the cwd-resolution branch is exercised
deterministically.

Add pytest-randomly so unit tests run in randomised order (they ran
single-process in deterministic alphabetical order, which hid ordering
bugs). Add a tests/unit autouse fixture that snapshots and restores the
Logger reaction/mode and the environment around every test, enforcing a
clean RAISE baseline so no test can leak global state forward. Raise the
coverage floor from 80% to 88% (currently 91.4%).

test_extinction_model_invalid (expects invalid input to fall back and
keep the current value) and Structures test_show_params (expects the
None-not-callable TypeError) only passed when an alphabetically-earlier
test had leaked Logger WARN mode. Set WARN explicitly in each so they
pass deterministically under randomised order.

Record issue 116 — adding mypy/pyright/ty would have caught the
wrong-arity _get_diagonal_shape call at lint time. Notes an incremental
rollout to manage the initial backlog.

Verify-review-1 finding 1: pytest-randomly was not in the plan's
pre-approval list. It was approved by direct user request in the
conversation (AGENTS.md §Architecture first approval path), so it need
not be removed; record that approval in the plan's Decisions section for
traceability.

On feature-branch pushes the tier selection is 'not pr and not
nightly'. Every integration test is auto-marked pr, so the selection is
empty and pytest exits 5 ('no tests collected'), failing the step under
set -euo pipefail. Add a run-pr-tier output from env-prepare and gate
both integration-test steps (source-test and package-test) on it, so
integration runs only in the pr tier (PRs + develop/master) as designed
and is cleanly skipped otherwise.

test_print_path_is_stringified hardcoded a POSIX path; str(Path) uses
backslashes on Windows. Compare against str(Path(...)) instead.

test_read_posterior_payload_reads_aux_arrays intermittently hit
FileNotFoundError creating an HDF5 file in a fresh tmp dir on Windows CI
(all sibling h5py tests share the pattern and passed, so this is an
environmental flake). Force the parent dir to exist and pass a str path
to remove the deterministic vectors.

The low-level sidecar helpers (_create_dataset, _read_dataset,
_read_payload_group, _write_payload_group, _read_posterior_payload) take
an open h5py handle and don't care whether it is disk-backed. Writing
real .h5 files to tmp_path dragged the platform filesystem into these
unit tests and caused an intermittent Windows ENOENT on file creation.

Add an _in_memory_h5() context manager using h5py's core driver
(backing_store=False) and route the 10 low-level tests through it: no
disk, no paths, hermetic, faster, and immune to FS flakiness on every
platform. Real on-disk tmp_path files remain only for the two
public-API tests that read a sidecar from a directory. Replaces the
mkdir+str band-aid; module coverage unchanged at 99%.

test_display_path_falls_back_to_absolute monkeypatched
pathlib.Path.relative_to to raise unconditionally, to simulate the
different-Windows-drive case (which POSIX cannot reach with walk_up).
That global patch also broke pytest's own relative_to call in its result
reporter (cwd_relative_nodeid -> bestrelpath), which only runs when the
invocation dir differs from the rootdir -- i.e. the package-test job
(pytest ../tests/unit from a subproject), crashing it with INTERNALERROR
'on a different drive'. Make the fake raise only for the resolved target
path and delegate every other call, so pytest internals keep working
while the except-ValueError fallback is still exercised cross-platform.