New versions of scripts for HyperPipe

[mebiri]

Updated files:
-util_ConstructIntrinsicPosterior_GenericCoordinates.py -> CIP_faster_v2.py
-allows n_events_to_analyze > 1
-conditional imports speed up initialization
-util_ConstructEOSPosterior.py -> util_ConstructHyperPosterior.py
-implements external rotation differently than rift_O4d AI code
-util_HyperparameterPuffball.py -> util_HyperPuffball_new.py
-implements external rotation, etc. differently than rift_O4d AI code
-Also 1 change to create_eos_posterior_pipeline to enable new argument structure for PUFF (arguments for parameter bounds now match CEP/CHP format param:[lower,upper])

[mebiri]

Actually that last change to create_eos_posterior_pipeline has not been done in this PR; line 191:
puff_args = puff_args.replace('[', ' \'[') -> puff_args = puff_args.replace('[', '\'[') (removed space inside quotes)

[oshaughnessy-junior]

Thanks for this. We audited the three files against current oshaughnessy-junior:rift_O4d (76ead4c) and ran small known-answer tests. Summary below. We plan to port the useful ideas into the existing tools rather than merge new parallel files; that port PR will be linked here.

Where you are right

• --n-events-to-analyze > 1. Our CIP accepts the flag but evaluates only one row. Under create_eos_posterior_pipeline chunking, (n-1)/n of the grid is skipped with no error. That is a real bug in our line.
• Startup cost. Conditional imports save about 6 s per job (--help ~3 s vs ~9 s on ldas-grid). The cost comes from corner (via arviz, 5.4 s) and senni (torch, 3.7 s), both of which were already unconditional in March.
• Coordinate passing. The hyperpipe never passes the coordinate module to the puff stage (RIFT/hyperpipe/coords.py, to_puff_args), so a rotated puff cannot be configured from a hyperpipe config. This is the main thing we will fix.
• param:[lo,hi] ranges for the puffball, and the one-character quoting fix in create_eos_posterior_pipeline (line 191; lines 138 and 216 have the same pattern).
• Comment bulk in CEP: the measurement table at L866-915 and long narrative comments belong in a design note.

Problems in the PR files

File	Problem	Evidence
CIP_faster_v2.py	Forked from CIP at d87a815 (March); drops 53 later commits, 29 of them already in this PR's base	git log d87a815f4..rift_O4d
	Every non-EOS run crashes: str(opts.using_eos_index+i) with index None (L2611)	measured, rc=1
	Per-file output names built from an already-mutated name: idx=0, n=4 writes MARG-0-{0,1,3,6}; neighbouring chunks overwrite rows	measured
	Lacks options that util_RIFT_pseudo_pipe, cip_pipeline.py, helper_LISA_Events and NMB (nmbseq:) emit
util_ConstructHyperPosterior.py	Gaussian at (3,3), 45 deg rotation, true lnZ = 19.33: CEP gives 19.34 in both plugin modes. CHP with physical-frame ranges gives mean (0.63, 2.78), lnZ = -26.7, with no warning	measured
	Re-introduces fixed bugs: double log of the likelihood (lnZ 2.96 vs 19.33); --internal-use-lnL with the adaptive sampler (lnZ 3.52); tempering_exp ignored	measured
	Loads the plugin only when --supplementary-coordinate-function is set, so under hyperpipe (which passes only -code) the plugin is dropped silently; .py paths and rift_default fail
util_HyperPuffball_new.py	CRLF line endings (python\r: No such file)	measured
	With no inverse available, turns rotation off but still applies rotated-frame bounds to physical coordinates
	--downselect-parameter x ... --reflect-parameter x (the old idiom, used in your hyperpipetests/Analysis/Makefile) now raises	measured
	The error-handling fix in 46231ee landed on the root-level copy that 1fb4aa2 deleted

For linear maps the two puffball rotation designs agree (covariance and reflection match). The CEP refusals that look like bloat are correctness fixes; the runs above show the wrong answers they prevent.

What the port will contain (all additive; no flag removed or renamed)

• CIP: lazy imports; working n>1 with the same output names as n separate jobs; --chunk-save; --save-hyperfile-only.
• CEP: optional get_bounds hook via --get-range-from-external and --external-range-args k=v, which raises on failure.
• Puffball: name:[lo,hi] ranges; reflection from a range without a downselect.
• Hyperpipe: the coordinate module reaches the puff stage.
• create_eos_posterior_pipeline quoting fix.

If your rotation plugin keeps physical column names, a small adapter that renames the outputs will let it run under CEP's plugin interface. We can help write it.

[oshaughnessy-junior]

Port PR: oshaughnessy-junior#383. It adds hyperpipe coordinate passing to the puff stage (post.coord-input-parameters, puff.coord-basis), name:[lo,hi] puff ranges, the get_bounds hook in CEP and the puffball, and CIP multi-row support with --chunk-save and lazy imports. Evidence is in its first comment.