test: re-baseline post-propagation in-air fraction after TauRunner decay fix

[thomwg11]

Context

Post-propagation in-air fraction (test/test_injection_regression.jl) has been failing on every PR branch since late July:

Evaluated: isapprox(0.24050632911392406, 0.2; atol = 0.02)

Identical to the digit across four CI runs on three branches — deterministic, not noise.

Root cause: upstream TauRunner, not TamboSim

icecube/TauRunner#72 (361661f, 2026-07-22) replaced the charged-lepton decay test in both SlabPropagator.jl and SphericalBodyPropagator.jl:

if particle.energy <= particle.mass * (1.0 + 1e-6)          # old — never fired
if Base.invokelatest(PROPOSAL_FN[:has_decay], secondaries)  # new

Taus decay in flight while still carrying nearly all their energy, so the old rest-mass test almost never fired: PROPOSAL decayed the tau, but TauRunner handed back a surviving tau at the decay vertex. Upstream's own regression test records 0/5 decays before, 5/5 after.

Downstream, this flips the branch at src/injection/inject.jl:265 — TauRunner now returns a ν_τ rather than a τ, so inject_neutrino_event no longer returns early and instead samples a vertex via force_interaction_vertex. That moves where PROPOSAL starts relative to the terrain, which is exactly what this test measures.

Last green main run was 2026-07-21, the day before the fix landed. It only bites CI because Manifest.toml is untracked (df53a9e) and Project.toml pins TauRunner by URL with no rev, so CI resolves upstream HEAD every run while local checkouts stay on whatever they last resolved.

Changes

The new TauRunner behavior is the physically correct one, so the regression values are re-baselined against it:

	old	measured	new
γ=1.0	0.200	0.2405 (95/395)	0.240
γ=2.0	0.197	0.2047 (79/386)	0.205

γ=2.0 had drifted for the same reason and was still inside the ±0.02 tolerance; re-baselined too rather than left to fail later. Tolerance unchanged.

Also adds an @info line printing the measured fraction and its numerator/denominator, so the next drift is readable straight off a CI log instead of needing a probe branch.

Verification

Measured on run 33823627332 (4151 pass / 1 fail — the single failure being this assertion), confirmed green on 33824911441: 4152 pass / 0 fail, both fractions reproducing exactly.

Note

TauRunner is deliberately left unpinned, per discussion. Anyone with a stale local Manifest.toml should Pkg.update("TauRunner") (in the repo and in examples/) — on the pre-fix tree the old 0.200 baseline is what reproduces.

🤖 Generated with Claude Code

[kcarloni]

Thanks for fixing this! I changed the wording because I often feel like Claude is too wordy in describing previous errors, which only adds unnecessary confusion.

[kcarloni]

Approved