[Jammy2211]

HowToGalaxy Chapter 3 previously jumped from Bayesian regularization straight to the combined MGE/pixelization search. It now explains why an evidence-preferred reconstruction can still leave residuals, then teaches brightness-guided mesh placement and adaptive smoothing before the combined model fit.

The five-lesson sequence is Pixelizations → Bayesian Regularization → Fit Problems → Adaptive Reconstruction → Model Fit. The evidence explanation uses the Goldilocks analogy and explicitly chooses the largest natural-log evidence. The new experiments use directly observed galaxies and their PSF/noise model. The original combined-model lesson's executable AST is unchanged; it moves from Tutorial 3 to Tutorial 5.

Scripts Changed

• tutorial_1_pixelizations.py: preserve the original walkthrough; add non-negativity, mapping-support and resolution caveats; complete the contents.
• tutorial_2_bayesian_regularization.py: clarify evidence selection, coefficient comparisons and feature reliability; complete the contents and onward links.
• tutorial_3_fit_problems.py: compare three increasingly compact galaxies with separately optimized constant regularization, motivating mesh resolution and spatially varying smoothing.
• tutorial_4_adaptive_reconstruction.py: combine Delaunay/Overlay, Hilbert sampling, ConstantSplit and AdaptSplitPower in one lesson, with bounded-memory evidence scans and explicit limits on interpretation.
• tutorial_5_model_fit.py: preserve the three original chained searches and add a bridge explaining the distinction between whole-galaxy adaptation and residual-component decomposition.

All scripts are under scripts/chapter_3_pixelizations/. Regenerated notebooks, READMEs and navigation/catalogues accompany the source. The two new compact-galaxy experiments declare full_datasets in their developer Env blocks.

Test Plan

• Independent review CLEAN at 5d98c61, with dispositions for the review surface's claims.
• All 30 generated notebooks validate; Chapter 3 narrative parity, exact contents lists and terminal sections pass. Navigation, catalogue determinism, API audit, strict environment declarations and search-memory checks pass.
• Full-size Tutorials 1, 3, 4 and 5 pass, including all new coefficient scans. The inherited full-size Tutorial 2 fine-mesh comparison was terminated with exit 143 after the revised evidence-selection section completed; this is recorded as incomplete, not a full-size pass. Its complete capped CI run passes.
• The nonlinear searches use PYAUTO_TEST_MODE=2: their likelihood and chaining paths execute, but posterior convergence is not assessed.
• Canonical repository smoke: all 30 executable scripts pass (zero failures, timeouts or skips). New Tutorials 3 and 4 execute at full size in 150.5s and 197.5s respectively. Heart’s smoke launcher has no HowToGalaxy entry, so the repository-owned CI runner was used in an isolated Python 3.12 environment with the declared optional dependencies.
• No unit-test directory exists in this teaching repository; script execution and generated-artifact checks are the relevant test surface.

Autonomous ship decision

The user explicitly requested autonomous adaptation. This docs task is supervised / judge: prepare the complete reviewable PR and leave merging to the human. The rejected alternative was stopping midway for repeated prose/ship approvals. To revert the complete source and generated change after merging, run git revert 5d98c61491cd7f03486731a855089e1b11d87d06.

Validation checklist (autonomous run)

• Effective level: supervised (large tutorial/docs adaptation); consequence: judge.
• Plan: recorded on issue docs: port Chapter 3 evidence and adaptive reconstruction teaching #101 and in the active Mind prompt.
• Gate: tests: generated/artifact/API checks pass (no pytest suite); smoke: 30/30 pass; independent review: CLEAN with claim dispositions; Heart: YELLOW, limited to the same-conversation acknowledged AutoFit workspace timeout. A missing current release rehearsal is recorded as a stale evidence gap; this is not a release.
• Human: plan sound in hindsight?
• Human: diff matches the requested Galaxy adaptation?
• Human: merge, amend or reject.

Closes #101.