fix: restore Julia 1.10 support for Double64 intlog2 path + re-enable downgrade CI

[ChrisRackauckas-Claude]

Ignore until reviewed by @ChrisRackauckas.

Problem

SciML policy requires every library to support Julia 1.10 (LTS), but the Core suite errored on 1.10 at latest compat and on master. The downgrade workflow was disabled (if: false, tracking #208).

The Issue-206 test "Double64 with pade order 13" errored on Julia 1.10:

exponential!(::Matrix{Double64}, ExpMethodGeneric{Val{13}()})
  -> intlog2(::Double64)
  -> ceil(UInt64, ::Double64)
  -> trunc(::Type{UInt64}, ::Double64)   # MethodError on Julia 1.10

Julia 1.11's Base added a generic trunc(::Type{<:Integer}, ::AbstractFloat) fallback; 1.10 lacks it, and DoubleFloats never defines trunc for Unsigned in any version. So exponential! errored on 1.10. Reproduced: 326 passed, 1 errored on Julia 1.10.

Fix (src/exp_generic.jl)

Route intlog2's integer-conversion path through a signed Int instead of UInt:

intlog2(x) = x > typemax(Int) ? ceil(Int, log2(x)) : intlog2(ceil(Int, x))

ceil(Int, x) only needs trunc(::Type{Int}, ::AbstractFloat), which DoubleFloats does define (it defines signed trunc for Int16/Int64/Int128/... but never Unsigned). The argument x is a positive operator norm bounded above by 2^62 (guarded just upstream), so the signed conversion is in-range and behavior is unchanged on 1.11+. Verified byte-for-byte equal to the old ceil(UInt, x) result across 10k random + boundary inputs (0 mismatches), including the 2^62 overflow guard.

Re-enable downgrade CI + corrected floors (Project.toml)

Removed if: false from .github/workflows/Downgrade.yml. Raised the lowest [compat] floors that resolve and pass the Core suite on Julia 1.10:

dep	old	new	reason
DoubleFloats	1	1.1.14	DoubleFloats 1.0.x pins GenericSchur to <=0.4, conflicting with this package's GenericSchur 0.5.3; 1.1.14 is the first DoubleFloats allowing GenericSchur 0.5
StaticArrays	1.2	1.9.8	versions < 1.9.8 either fail to precompile on Julia 1.10 (require_one_based_indexing not defined) or hit an ambiguous lu(::StaticMatrix, ::Val{true}) against LinearAlgebra's deprecated method

(GenericSchur 0.5.3 and all other floors were left unchanged — they resolve fine.)

julia = "1.10" and the workflow's julia-version: "1.10" are unchanged, per the LTS policy.

Local test results (Julia 1.10.11)

Latest compat:

Quality Assurance | 10  10
Basic Tests       | 329  1(broken)  330
Testing ExponentialUtilities tests passed

Downgrade floors (DoubleFloats 1.1.14, GenericSchur 0.5.3, StaticArrays 1.9.8):

Quality Assurance | 10  10
Basic Tests       | 329  1(broken)  330
Testing ExponentialUtilities tests passed

The 1 broken is a pre-existing @test_broken (kiops, test/basictests.jl:293), untouched here. The previously-erroring Issue-206 Double64 test now passes.

🤖 Generated with Claude Code

[ChrisRackauckas-Claude]

Update: downgrade made strict (allow-reresolve: false)

Switched the Downgrade.yml caller to the strict SciML default allow-reresolve: false (was true). This forces Pkg.test to instantiate the exact deps-floor + latest-transitive set with no reconciliation, which is the rigorous downgrade check.

Verified locally on Julia 1.10.11, deps-mode (only [deps] [compat] pinned to floors), Pkg.test(; allow_reresolve=false), GROUP=Core:

Quality Assurance | 10  10
Basic Tests       | 329  1(broken)  330
Testing ExponentialUtilities tests passed

No floor raises were needed — the current [deps] floors (Adapt 3.4.0, ArrayInterface 7.1, GPUArraysCore 0.1, GenericSchur 0.5.3, PrecompileTools 1) are already strict-consistent with the latest transitive ecosystem on Julia 1.10. The 1 broken is the same pre-existing @test_broken (kiops, test/basictests.jl:293).

Ignore until reviewed by @ChrisRackauckas.