Realias result on the DiffResults paths (fixes #269 and #251)

[devmotion]

Fixes #269. Fixes #251.

DiffResults requires the object returned by a mutating call to be used; we returned the argument binding instead. Two bugs follow from that:

• Value is sometimes not set when using DiffResults #269: on Julia 1.8 – 1.11.4 the optimizer forwards a stale pre-call read of result's fields across the call that writes them (1.10 LTS (and 1.11.0–1.11.4): SROA deletes field writes to a mutable that escapes to a non-inlined callee (silent wrong results; fixed by #57201 in 1.11.5) JuliaLang/julia#62812, fixed in 1.11.5, never backported to the 1.10 LTS).
• DiffResults objects are not re-aliased properly #251: for an ImmutableDiffResult — what DiffResults.GradientResult(x) returns for a StaticArray — value! cannot mutate and returns a new struct, so the primal value is dropped on every Julia version. Only the wrapper is immutable, so the derivatives still come out correct and only DiffResults.value goes stale.

The rebind has to be unbroken: gradients.jl, tape.jl and utils.jl each still fail on their own. utils.jl is the load-bearing one — its leaf extract_result!/extract_result_value! methods already realias, but every caller discarded the result.

Two changes beyond the mechanical rebind:

• The Tuple methods in utils.jl looped over result[i], which is not assignable, so a returned struct had nowhere to go. They now map over the tuple, matching construct_result(input::Tuple) just below them. This also removes the pattern that allowed the bug: map has no place to drop a return value. Since map over tuples silently truncates to the shorter argument, the two methods taking both a result and an input tuple now constrain them to a common NTuple{N,Any}, so a mismatched arity is rejected by dispatch rather than silently ignored.
• hessian!(result::DiffResult, tape, input) builds an inner DiffResult holding the gradient in its value slot. MutableDiffResult copies into the shared buffer so it propagates, ImmutableDiffResult rebuilds the struct instead, so the gradient was dropped. That path now reads the gradient back off the inner result.

It costs nothing: allocations are byte-identical across all 26 gradient/jacobian/hessian entry points measured A/B, array and tuple, and inference stays concrete throughout. The compiled-tape tuple paths still allocate 0 — the rebuilt tuple is concretely typed and doesn't escape, so it is elided.

New tests cover gradient!, jacobian! and hessian! with MVector inputs via the config path, a tape and a compiled tape, plus a tuple of immutable results. The #269 test was verified to fail on Julia 1.10 with either half of the fix reverted, under the same --check-bounds=yes --code-coverage=user flags CI uses. Full suite green on 1.10, 1.11 and 1.12.

Rebased onto master, which now carries #292 and the StaticArrays = "1.6.4" bound this needs. The #251 Hessian test goes through HessianConfig(result, input), which builds a tracked buffer with similar(output, TrackedReal{D,D,Nothing}); TrackedReal is not isbitstype, and StaticArrays only returns a writable SizedArray for non-isbits eltypes from 1.6.4 on. That is what the Julia min job hit on Julia 1.0, where the resolver picked StaticArrays 1.3.6; that job no longer exists.

The redundant v1.17.3 bump is dropped, since master released it, and the version is set to v1.18.0 for tagging.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 86.08%. Comparing base (2b1bbb5) to head (b69f5d7).

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #295      +/-   ##
==========================================
+ Coverage   85.53%   86.08%   +0.54%     
==========================================
  Files          19       19              
  Lines        1964     1947      -17     
==========================================
- Hits         1680     1676       -4     
+ Misses        284      271      -13     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.