Gibbs test | Fix dynamic model test in Gibbs sampler suite

[AoifeHughes]

The existing dynamic model test was problematic:

• Used a complex Dirichlet Process mixture model that was slow (10,000 samples)
• Only performed regression testing against "old" values with a TODO indicating it wasn't ideal
• Failed to properly test the core dynamic model functionality that Gibbs sampling should support

Here I've changed this with:

a new dynamic_bernoulli_normal model that:

• Has varying dimensions: A Bernoulli variable b controls whether the model has 1 or 2 parameters
• Has analytical ground truth: Marginal likelihoods and posterior for b can be computed exactly
• Is fast to sample: Only 1,000 samples vs 10,000, runs in ~7-12 seconds
• Tests core functionality: Validates that Gibbs sampling correctly handles dimension changes without errors

[AoifeHughes]

Leave this as a draft for a teeny bit, I need to properly test and fully understand how julia tests work so I'm not breaking what I think I understand - no need for any reviews yet.

[AoifeHughes]

This should address: #2402 when completed.

[mhauru]

As discussed on Slack, feel free to change the base branch of this to main. Since it's not a user-interface breaking change no need to make a breaking release for it and make it go through the breaking branch.

[AoifeHughes]

Results from my local testing:

Test Summary:                                          | Pass  Total      Time
Turing                                                 | 2613   2613  19m50.9s
  Aqua                                                 |           0      0.1s
  AD                                                   |           0      0.0s
  essential                                            |           0      0.0s
  samplers (without AD)                                |           0      0.0s
  inference with samplers                              | 2613   2613  19m50.8s
    GibbsContext                                       |  868    868     27.7s
      type stability                                   |  868    868     27.7s
    Invalid Gibbs constructor                          |    8      8      0.1s
    Sampler call order                                 |    1      1     53.9s
    Equivalence of RepeatSampler and repeating Sampler |    1      1      4.5s
    Gibbs warmup                                       |    8      8      2.3s
    Testing gibbs.jl                                   | 1727   1727  18m19.3s
      Gibbs constructors                               |   13     13     19.7s
      Gibbs inference                                  |   22     22   6m06.7s
      transitions                                      |           0      2.1s
      dynamic model with analytical posterior          |    5      5      8.4s
      dynamic model with ESS                           |    6      6     33.4s
      dynamic model with dot tilde                     |           0     11.3s
      Demo model                                       | 1564   1564   8m32.2s
      multiple varnames                                |   20     20      1.5s
      non-identity varnames                            |    2      2      9.2s
      submodels                                        |    2      2     11.3s
      CSMC + ESS                                       |   42     42     38.9s
      CSMC + ESS (usage of implicit varname)           |   42     42     51.8s
      externalsampler                                  |    8      8     28.2s
      linking changes dimension                        |    1      1      4.5s
  variational algorithms                               |           0      0.0s
  mode estimation                                      |           0      0.0s
  variational optimisers                               |           0      0.0s
  stdlib                                               |           0      0.0s
  utilities                                            |           0      0.0s
───────────────────────────────────────────────────────────
                                Time          Allocations
                          ───────────────   ───────────────
     Total measured:           1191s             534GiB

Section           ncalls     time    %tot     alloc    %tot
───────────────────────────────────────────────────────────
inference              1    1191s  100.0%    534GiB  100.0%
  mcmc/gibbs.jl        1    1191s  100.0%    534GiB  100.0%
───────────────────────────────────────────────────────────     Testing Turing tests passed

[github-actions]

Turing.jl documentation for PR #2579 is available at:
https://TuringLang.github.io/Turing.jl/previews/PR2579/

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 85.58%. Comparing base (7ebde76) to head (771934c).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2579   +/-   ##
=======================================
  Coverage   85.57%   85.58%           
=======================================
  Files          22       22           
  Lines        1456     1457    +1     
=======================================
+ Hits         1246     1247    +1     
  Misses        210      210           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[coveralls]

Pull Request Test Coverage Report for Build 15822411978

Details

• 2 of 2 (100.0%) changed or added relevant lines in 1 file are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage increased (+0.01%) to 85.646%

---

Totals
Change from base Build 15765001949:	0.01%
Covered Lines:	1247
Relevant Lines:	1456

---

💛 - Coveralls

[Copilot]

Pull Request Overview

This pull request refactors several model definitions and tests to adopt the new default-type parameter syntax and introduces a new dynamic model (dynamic_bernoulli_normal) with an analytical posterior for faster and more robust testing of the Gibbs sampler. Key changes include:

• Updating function signatures across multiple tests to use the new parameter syntax with parentheses.
• Introducing a dynamic model with a Bernoulli variable controlling parameter dimensionality.
• Simplifying and speeding up tests while ensuring core Gibbs sampling functionality is exercised.

Reviewed Changes

Copilot reviewed 4 out of 5 changed files in this pull request and generated no comments.

File	Description
test/mcmc/hmc.jl	Updated the default type parameter syntax in model definitions.
test/mcmc/gibbs.jl	Updated function signatures and added a new dynamic model test.
test/mcmc/Inference.jl	Applied similar parameter signature updates across model tests.
src/mcmc/external_sampler.jl	Updated constructor and helper function parameter syntax.

Comments suppressed due to low confidence (2)

test/mcmc/gibbs.jl:564

• [nitpick] The check for θ[2] values only verifies that one of the conditions is met. Consider splitting the test into two separate assertions to clearly validate that when b=0, θ[2] is missing, and when b=1, θ[2] exists.

            @test n_missing_theta2 > 0 || n_present_theta2 > 0  # One of these should be true

src/mcmc/external_sampler.jl:47

• [nitpick] Consider adding a comment or type annotations to clarify the expected type and meaning of the 'unconstrained' flag, which would improve maintainability and readability of the code.

function requires_unconstrained_space(

[yebai]

Thanks @AoifeHughes -- good work!

[penelopeysm]

Minor, but it might be better to use LogExpFunctions.logsumexp here rather than rolling our own version.

[yebai]

yes

[penelopeysm]

Otherwise, if @yebai if happy, I am too