Fix Jacobian for dynamic-to-fixed transforms

[yebai]

Return target forward Jacobians, add a DynamicLink-to-FixedTransform regression test

Fix #1407

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 81.84%. Comparing base (a8b0145) to head (47431f5).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1442      +/-   ##
==========================================
+ Coverage   81.70%   81.84%   +0.13%     
==========================================
  Files          50       50              
  Lines        3581     3581              
==========================================
+ Hits         2926     2931       +5     
+ Misses        655      650       -5     

☔ View full report in Codecov by Harness.
📢 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.

[github-actions]

DynamicPPL.jl documentation for PR #1442 is available at:
https://TuringLang.github.io/DynamicPPL.jl/previews/PR1442/

[shravanngoswamii]

Can we say inv_logjac is now dead in the FixedTransform? -- needs some docs update, is it fine if I push few commits to your PR?

[github-actions]

Benchmarks @ 47431f5

Performance Ratio: gradient time divided by log-density time.

For very small models these ratios are noisy across runs and machines; raw primal and gradient timings are more reliable. The benchmarks are aimed at DynamicPPL developers and mainly catch obvious allocation or type-stability regressions. See benchmark notes for details.

===================================================================================================
                                               eval                       gradient                 
                                            ----------  -------------------------------------------
Model                        dim    linked      primal     FwdDiff    RvsDiff    Mooncake    Enzyme
---------------------------------------------------------------------------------------------------
Simple assume observe*         1     false     4.63 ns       12.73    1471.34       40.50     12.38
Simple assume observe*         1      true     4.63 ns       12.54    1716.71       40.18     12.14
Smorgasbord                  201     false      6.0 μs       66.30     133.52        6.93      9.71
Smorgasbord                  201      true     7.59 μs       76.12     143.70        6.35      7.03
Loop univariate 1k          1000     false     17.5 μs      982.79     307.90        8.26      6.66
Loop univariate 1k          1000      true     19.0 μs     1483.43     286.34        7.64      6.18
Multivariate 1k             1000     false     21.9 μs      355.34      79.18        9.55      2.79
Multivariate 1k             1000      true     27.1 μs      267.36      56.16        8.12      2.97
Loop univariate 10k        10000     false    172.0 μs    12079.17     329.25        8.55      6.72
Loop univariate 10k        10000      true    186.0 μs    11879.77     308.99        7.91      6.17
Multivariate 10k           10000     false    194.0 μs     5747.98      91.95       11.64      2.30
Multivariate 10k           10000      true    197.0 μs     5688.03      90.05       11.42      2.29
Dynamic                       15     false     1.36 μs         err      44.49       14.89     12.02
Dynamic                       10      true     1.92 μs        1.79      61.00       22.21     20.36
Submodel*                      1     false     4.64 ns       12.60    1749.19       40.41     12.38
Submodel*                      1      true     4.64 ns       12.60    1882.50       40.46     12.26
LDA                           12      true     23.8 μs        0.45       1.94       33.51       err
===================================================================================================

Main @ a8b0145

===================================================================================================
                                               eval                       gradient                 
                                            ----------  -------------------------------------------
Model                        dim    linked      primal     FwdDiff    RvsDiff    Mooncake    Enzyme
---------------------------------------------------------------------------------------------------
Simple assume observe*         1     false     4.63 ns       12.63    1561.33       39.65     12.43
Simple assume observe*         1      true     4.63 ns       12.64    1718.38       39.92     12.30
Smorgasbord                  201     false      6.0 μs       66.66     133.34        6.94      9.46
Smorgasbord                  201      true     7.61 μs       66.45     152.95        6.21      6.98
Loop univariate 1k          1000     false     17.6 μs      942.60     310.17        8.43      6.73
Loop univariate 1k          1000      true     19.0 μs     1461.31     284.67        7.66      6.14
Multivariate 1k             1000     false     22.1 μs      355.85      79.90        8.83      2.96
Multivariate 1k             1000      true     32.0 μs      332.81      51.81        7.18      3.30
Loop univariate 10k        10000     false    173.0 μs    12519.57     353.50        8.67      6.88
Loop univariate 10k        10000      true    186.0 μs    12732.46     322.63        8.34      6.36
Multivariate 10k           10000     false    205.0 μs     5161.69      85.75       11.73      2.41
Multivariate 10k           10000      true    206.0 μs     5102.01      89.43       11.27      2.57
Dynamic                       15     false     1.41 μs         err      43.50       14.75     11.69
Dynamic                       10      true     1.97 μs        1.88      62.59       18.36     24.82
Submodel*                      1     false     4.63 ns       12.64    1743.24       39.98     12.40
Submodel*                      1      true     4.63 ns       12.69    1900.02       40.27     12.34
LDA                           12      true     23.3 μs        0.63       2.06       34.08       err
===================================================================================================

Environment

Julia Version 1.11.9
Commit 53a02c0720c (2026-02-06 00:27 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: 4 × AMD EPYC 7763 64-Core Processor
  WORD_SIZE: 64
  LLVM: libLLVM-16.0.6 (ORCJIT, znver3)
Threads: 1 default, 0 interactive, 1 GC (on 4 virtual cores)

[shravanngoswamii]

@yebai Ready to merge, I will release it if you allow merging!

[yebai]

It's better to have a look from a fresh perspective. @sunxd3

[sunxd3]

Sorry it took me a while, chasing a ghost that's turned out to be nothing. The changes here look good to me. Thanks for working on this!