Implement observed asymptotic CLs likelihood limit

[davidwalter2]

Problem

The observed likelihood-based asymptotic CLs upper limit was never produced:

• The call in rabbit_limit.py was commented out (# TODO: make it work) and the channel/mapping case was a pass (# TODO: implement for channels); the likelihood_asymptoticLimits_observed output histogram was commented out — so the bsm CI jobs silently yielded NaN.
• The old compute_likelihood_limit had a real Jacobian bug: it differentiated sqrt(2*NLL) instead of sqrt(2*(NLL - nllvalreduced)), mis-scaling the constraint Jacobian fed to the joint trust-constr solve.
• The modified statistic q̃_μ for a negative best-fit signal strength was a TODO.

Changes

Rewrite compute_likelihood_limit to reuse the same profiled-NLL machinery as the (already working) expected limit, instead of the bespoke two-fitter joint solve: evaluate the profile-likelihood test statistic with the data fit (q_μ) and the background-only Asimov fit (q_{μ,A}) and root-find μ over CLs(μ) = cl_s, using the same algebra as the gaussian path (Cowan, Cranmer, Gross, Vitells, arXiv:1007.1727).

• rabbit/fitter.py: add Fitter.profiled_nll_at_poi — fix the POI, profile all other parameters (freeze_params/minimize/reduced_nll), restore state. The literal profile-likelihood test-statistic definition.
• rabbit/asymptotic_limits.py: replace the buggy bespoke solver with the CLs root-find; implement the modified statistic q̃_μ (Eq. 16) for μ̂ < 0, consistent with the gaussian xobs < 0 branch; handle the channel/mapping case by reporting the mapped observable at the limit point (same convention as the expected channel limit).
• bin/rabbit_limit.py: wire the observed likelihood limit for both the parameter and channel cases; enable the likelihood_asymptoticLimits_observed output.
• .github/workflows/main.yml: assert in both bsm CI jobs that the observed likelihood limit exists, is finite/positive, and sits within the expected ±2σ band (previously silently NaN).

Verification

End-to-end on the existing bsm CI flow (tests/make_tensor.py --bsm):

case	gaussian obs	likelihood obs	likelihood expected (median / ±2σ)
parameter bsm1	1.89	2.06	2.30 / [1.60, 3.78]
channel bsm1_sig_mixing	0.043	0.034	0.046 / [0.028, 0.091]

Each method's observed limit sits just below its own expected median (consistent with the slightly negative μ̂; modified statistic engaged). isort/black/flake8/pylint clean, workflow YAML valid.

🤖 Generated with Claude Code