Fix: audit logscale missing

[docxology]

[Verified audit fix] — branch fix/audit-logscale-missing.

Commit: Skip post-rule annotation processors when rule short-circuits to missing

Changes

src/message.jl | 5 ++++-
src/rules/gcv/w.jl | 6 +++---
2 files changed, 7 insertions(+), 4 deletions(-)

Verification

Confirmed against main in the ReactiveBayes audit (Julia 1.12.6); sources verified read-only. See the branch diff.

[bvdmitri]

Maintainer review — the defect is real and the one-line fix is correct. I have rebased this branch and pushed a regression test plus a CHANGELOG entry (thanks for enabling maintainer edits).

Rebased to a single concern

As submitted, this branch sat on top of #641's GCV commit, so its diff included src/rules/gcv/w.jl — an unrelated change. (#641 ⊂ #640 ⊂ #639 was the full stack; merging #639 would have landed all three at once.) Rebased onto main, so it now contains only the annotation change plus its test.

Verified reachable

Confirmed by code trace: MessageMapping sets result = missing and skips the rule when any inbound message or marginal is missing (src/message.jl:657-663) — the mechanism behind new_observation!(datavar, missing) — then runs post_rule_annotations! regardless. LogScaleAnnotations.post_rule_annotations! only defaults :logscale = 0 when all inputs are PointMass (logscale.jl:47-49); a missing message is not, so it falls to error(...). A legitimately deferred message became a hard crash, defeating the short-circuit.

One correction to the reasoning, in the fix's favour

I initially expected this fix to be insufficient — a missing-data message still reaches compute_product_of_two_messages, which calls post_product_annotations! unconditionally (message.jl:236), and LogScaleAnnotations.post_product_annotations! reads getlogscale on both sides (logscale.jl:68-69). That looked like it would simply trade the error for a KeyError(:logscale) one layer down.

It does not, and the reason is worth recording: the generic post_product_annotations! already dispatches on ::Missing for left_dist/right_dist (src/annotations.jl:168-190) and copies the other side's annotations through without ever consulting the per-processor method. So LogScaleAnnotations is never reached with an absent :logscale. The fix is sufficient as written — but that was load-bearing behaviour nobody had pinned, so the test now covers it explicitly in both orderings (missing-on-left and missing-on-right).

Regression test added

test/annotations/logscale_tests.jl, three testsets deliberately arranged so the important one cannot pass vacuously:

1. Control: a concrete non-PointMass message still reaches the processor and still errors. Without this, the missing test would also pass if the processor had simply stopped running altogether.
2. The regression: a missing message comes back as missing and carries no :logscale — no log-scale is invented for a message that never went through a rule.
3. Downstream: the deferred message multiplied against a concrete one, both orderings, asserting the concrete side's log-scale propagates.

src/message.jl	result
reverted to pre-fix main	2 errors (testsets 2 and 3)
with this PR's fix	5 pass

The control testset passes in both runs.

Note on the submitted description

The PR body listed src/rules/gcv/w.jl under "Changes" and said "Confirmed against main in the ReactiveBayes audit ... sources verified read-only", but carried no test. For a crash on a documented code path in the annotations layer, a test is the deliverable — now added.