From 3107b29b5b9e935dd156743c85479e65c27e82be Mon Sep 17 00:00:00 2001 From: Claude Date: Wed, 5 Aug 2026 08:21:38 +0000 Subject: [PATCH] chore: park tutorial_5_expectation_propagation as NEEDS_FIX MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit `LinearRegressionAnalysis.log_likelihood_function` returns a constant `-1` and ignores `instance`, so every initial sample drawn for that factor has an identical figure of merit and its per-factor search cannot initialise. At real sampling the `linear_regression` factor never completes a single update. The tutorial "passes" CI today only because HowToFit runs at `PYAUTO_TEST_MODE=2`, which bypasses the sampler and therefore the initializer — so the breakage is invisible to the gate while being the first thing a reader running the tutorial for real would hit. Parked rather than fixed: a real fix needs a likelihood for the m/c regression over `fwhm_list` (computed just above the class and then never used), which is tutorial authoring rather than a mechanical repair, and the tutorial gives no indication what the regression is meant to be against. NEEDS_FIX rather than a plain skip, so every mega-run surfaces it instead of it becoming a permanent silent exclusion. Surfaced by PyAutoFit#1454, which made the failure legible by naming the stale factor instead of dying on a raw InitializerException. Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_01GSBb1i58cNhTx64yxBGPfs --- config/build/no_run.yaml | 2 ++ 1 file changed, 2 insertions(+) diff --git a/config/build/no_run.yaml b/config/build/no_run.yaml index f959009..ab92eef 100644 --- a/config/build/no_run.yaml +++ b/config/build/no_run.yaml @@ -9,3 +9,5 @@ # are broken and parked as a to-do list. These are NOT permanent skips — # every mega-run surfaces them with a loud warning banner. Investigate the # failure, fix the underlying bug, and remove the NEEDS_FIX marker. + +- tutorial_5_expectation_propagation # NEEDS_FIX 2026-08-05 - LinearRegressionAnalysis.log_likelihood_function returns a constant -1, ignoring `instance`, so every initial sample has an identical figure of merit and the per-factor search cannot initialise. The tutorial only "passes" today because CI runs at PYAUTO_TEST_MODE=2, which bypasses the sampler; at real sampling the linear_regression factor never completes a single update. Fix needs a real likelihood for the m/c regression over fwhm_list (currently computed and unused), which is tutorial authoring, not a mechanical repair. See PyAutoFit#1454.