JET typo-mode: 36 reports on Julia 1.12 (1 real Euler-interp bug + ~35 correlated-branch may-be-undefined)

JET report_package(; mode = :typo) findings on Julia 1.12 (JET 0.11.x)

When the QA group runs the standard SciML JET typo check
(JET.test_package(SimpleDiffEq; target_modules = (SimpleDiffEq,), mode = :typo),
now enabled via run_qa in SciMLTesting 1.6) it is clean on Julia 1.10/1.11
(JET 0.9.x) but reports 36 possible errors on Julia 1.12 (JET 0.11.x). The
source is byte-identical to master, so these predate the QA-conversion PR
(#122) — the conversion merely turned the JET check on.

Reproduce

# in test/qa (SciMLTesting 1.6.0, JET resolves to 0.11.x on Julia 1.12)
using SimpleDiffEq, JET
rep = JET.report_package(SimpleDiffEq; target_modules = (SimpleDiffEq,), mode = :typo)
length(JET.get_reports(rep))   # 0 on 1.10/1.11, 36 on 1.12

The 36 findings, by category

1. One genuine latent bug (local variable u is not defined)
src/euler/euler.jl:199, in the in-place branch of the SimpleEulerIntegrator
interpolant:

else
    for i in 1:length(u)             # reads `u` before it is ever assigned
        u = @. (1 - Θ) * y₀ + Θ * y₁
    end
    return u
end

u is only bound inside the loop body, but length(u) reads it first — this
path throws UndefVarError if exercised. (Likely length(u) should be
length(y₁), and the in-place branch should write into a preallocated buffer
rather than rebind u.)

2. ~35 may be undefined reports on correlated-branch control flow
These are correct code that JET 0.9 could prove safe but JET 0.11 cannot,
because the guard that guarantees the binding lives in a separate if block:

• q11 in the adaptive step! error-estimate blocks
  (src/tsit5/atsit5.jl:293,378,546,676, src/tsit5/gpuatsit5.jl:278,
  src/verner/gpuvern7.jl:407, src/verner/gpuvern9.jl:575): q11 is bound only
  in the else of if iszero(EEst), then used under if EEst > 1. When
  EEst > 1, EEst is nonzero, so the else ran and q11 is bound.
• cur_t in the saveat save loops
  (src/tsit5/atsit5.jl:150,152,154,156, src/tsit5/gpuatsit5.jl:110,111,114,120,296,297,300,306,
  src/verner/gpuvern7.jl:138,139,207,216,425,426,499,508,
  src/verner/gpuvern9.jl:185,186,316,327,593,594,739,750): cur_t is bound only
  in the else of if saveat === nothing, then used under elseif saveat !== nothing.

Mitigation in place

test/qa/qa.jl marks the JET check jet_broken on Julia 1.12+ only
(jet_broken = VERSION >= v"1.12"), so the QA lane records Broken (green) while
these stand, and the still-clean 1.10/1.11 lanes keep running the hard
test_package check. The marker auto-flips to an Unexpected Pass (error) once
the reports are resolved, prompting removal of the gate.

Resolution

1. Fix the in-place Euler interpolant (euler.jl:199) — a real bug.
2. Initialize q11 / cur_t unconditionally (or restructure to a single
   if/else) so JET 0.11 can prove the binding, then drop the jet_broken gate.

Start by reproducing the Julia 1.12 report with JET.test_package in test/qa, then inspect test/qa/qa.jl and the cited interpolant and solver files. Address the undefined Euler path and the q11/cur_t findings across the listed files, rerun the QA checks on Julia 1.10–1.12, and confirm the jet_broken gate can be removed.