Make the repo the source for the analysis scripts

[marinerhemant]

scripts/pipeline/analysis/ had forked. The versions actually used to process the July 2026
beam time (13 scans, ~150k frames, four machines) lived outside the repo and had gained a
scan registry, environment-driven paths and several new tools. The repo copies were stale —
parentbeta_validate 133 vs 153 lines, parentbeta_reconstruct 214 vs 218 — so anyone
running from the repo got the old code.

This makes the repo the source.

Updated

• parentbeta_validate.py — scan registry + env-driven paths; runs on any host, any scan
• parentbeta_reconstruct.py — output prefix parameterised

Added

All driven by LAUE_WORK / LAUE_SCAN_DATA / LAUE_SCAN_{ALPHA,BETA} / LAUE_OUT_PREFIX:

script	purpose
run_analysis_chain.sh	the whole sequence, one command per scan
null_model.py	measures the random-orientation null on that scan
empirical_gate.py	re-scores the validated set against that measured null
exclusion_null.py	measured null for the β α-exclusion test
anchor_null.py	tests whether the retained-β anchor beats chance
variant_coherence.py	spatial coherence of the variant map vs a shuffle null
regrain.py	grains as contiguous regions, not orientation-only clusters
tolerance_sensitivity.py	how counts move with the clustering tolerance
big_grain_diagnostic.py, big_grain_split_test.py	is one large "grain" actually several?
collect_scan_metrics.py	cross-scan summary JSON
scan_map.py	unvalidated catalog map (was hardcoded to a single scan)
catalog_figures.py, validated_figures.py	report plates

Why these matter, not just that they exist

Three findings from this beam time are encoded here, each of which changed a number we would
otherwise have published:

1. The analytic p<1e-4 gate under-rejects. It assumes uniformly scattered peaks; real
   Laue peak fields are clustered, so the true null has much heavier tails. Across nine scans
   the measured α null maximum ranged 14–17 and β 11–16 — one inherited value misstates the
   others. On one dataset this moved the defensible β count by two orders of magnitude.
2. A grain is a contiguous region of consistent orientation. Orientation-only clustering
   merges spatially disjoint regions; the largest "grain" in one scan proved to be two
   crystallites 0.44° apart (73σ past a label-shuffle null). Splitting on connectivity moved
   α 325→614 and β 40→27 — opposite directions, which is what shows it is a definitional fix
   rather than a tuning knob.
3. A corroborating statistic must itself be tested. The retained-β anchor is compelling
   with few β clusters and worthless with many: against 2,537 candidates a random
   orientation lands within 1.74° of one 9% of the time.

Naming

The id6_* prefix named one scan; these tools are scan-agnostic, so they now follow the
directory's existing convention (null_model.py, not id6_null_model.py). The id6_10x10
registry key is kept for backward compatibility with existing outputs and carries a comment
noting it is specimen ID26 at 20×20 µm — not ID6 at 10×10. The folder name is wrong;
seven of these scans are mislabelled the same way.

README

Adds the chain documentation, the three rules above, and a gotchas section: tcsh noclobber
silently refusing writes, ls | wc -l returning 0 past ARG_MAX, ls -dt matching
.launch.log, CUDA_DEVICE_ORDER, pgrep -f matching its own command line, outputs
appearing in one late batch, and blank frames legitimately producing no output.

Depends on nothing in #5, but both come from the same campaign.