Provenance live-probe guard checks no R packages while its consumer requires four, so the testset errors on any bare checkout #43

Measured

The full Julia suite errors locally with:

live probes of what is actually installed   |  0 pass  1 error   3.8s
provenance: could not prove 'r': package 'dada2' is not installed in the R library in use

Zero passes before the error means it threw before the first @test.

The cause: the guard asks a different question than its consumer

test/unit/test_provenance.jl:439-448

r_ready = try
    PV.probe_r(; packages = String[], timeout = 30)   # <- probes NO packages
    true
catch err
    @info "Provenance: skipping live R probe" reason=sprint(showerror, err)
    false
end

if r_ready
    record = PV.probe_r(; timeout = 120)              # <- probes FOUR packages

src/core/provenance.jl:292

const R_PACKAGES = ["dada2", "Biostrings", "ShortRead", "vegan"]

The guard asks "can R be reached at all?". The consumer asks "can R be reached
and are dada2, Biostrings, ShortRead and vegan all installed?". On any checkout
where R exists but the packages do not — which is every bare checkout, and this
development box — the guard returns true and the consumer throws outside any
try, erroring the testset.

The @info "skipping live R probe" branch therefore cannot fire for the case it was
written for. The file's own header states the intent it fails to meet:

Nothing here shells out to a tool or evaluates R unless the tool or the package is
genuinely present, so the suite runs on a bare checkout.

Measured context

• All six external tools are absent locally (cutadapt, fastqc, multiqc,
  vsearch, swarm, cd-hit-est), so the tool loop above contributes nothing — the
  R branch is the sole error.
• renv reports "One or more packages recorded in the lockfile are not installed" on
  every julia start in this checkout, so the condition is the normal local state.
• This is adjacent to CI installs four of six pipeline tools: FastQC and MultiQC are absent, so prefilter QC is never exercised #30 but is a different defect: CI installs four of six pipeline tools: FastQC and MultiQC are absent, so prefilter QC is never exercised #30 is about FastQC/MultiQC never
  being installed in CI, this is about the R guard being unable to protect its own
  consumer.

Acceptance criteria

1. The guard probes exactly what the consumer will probe — pass the same package
   list to both, or probe once and reuse the record. A test asserting the two call
   sites agree is preferable to a comment saying they should.
2. On a checkout where R is present but R_PACKAGES are not, the testset skips
   loudly and passes — it must not error.
3. A skip must name which packages were missing, not merely that a probe failed.
   Per the standing rule that a skip is not a pass, the skip must be visible in the
   run summary rather than only in @info output that nobody reads.
4. Add the negative control: a case where R is reachable and the packages are absent
   must be distinguishable in the test output from a case where R is unreachable.

🤖 Generated with Claude Code

https://claude.ai/code/session_01X3hgXxWm6umMgZkjYyHnnm