meta(ad): Enzyme coverage — operator pullbacks, packed/GPU parity, Julia 1.13 #82

Umbrella for the open Enzyme-rule and Enzyme-coverage issues. The exchange and BC sweeps have declared-transpose rules over raw storage (_exchange_storage! / _bc_storage! + BlockLayout); the operator sweeps do not, one crash is fixed but not root-caused, none of it is tested on a GPU or a packed field, and four gradient tests are skipped on Julia 1.13.

Sub-issues

1. feat(ad): linear-operator pullback rules (DESIGN §6) — re-verify against current Enzyme #41 — linear-operator pullback rules (DESIGN §6). The upstream tracker it cited closed in February; first step is to re-run the mixed pointer/isbits MWE on current Enzyme and either unblock or re-point.
2. test(ad): GPU/packed-field AD parity for the exchange rules #43 — packed-field and GPU AD parity for the exchange rules. Decide and document whether GPU reverse mode is supported or refused; today it is untested.
3. test(ad): re-enable Enzyme gradient tests skipped on Julia 1.13 #108 — re-enable the Enzyme gradient tests skipped on Julia 1.13 (test/autodiff.jl, test/enzyme_rules.jl). CI runs only 1.10 and 1.13, so the dκ parameter-gradient guard now runs on half the matrix.

Optional, no issue: the Linux x86_64 segfault fixed in 9fc1b02 (formerly #40) was never root-caused. That commit removed the single-grid apply_bc!/fold_bc! rule and made the remaining rule bodies allocation-free in one step, so the candidates are allocation in the rule body and the removed rule's mixed pointer/isbits AbstractGrid argument (same class as EnzymeAD/Enzyme.jl#2707). Re-bisect with an upstream MWE only if a single-grid rule is ever wanted again.

Invariants that bind every sub-issue

• A rule is an exact substitution of a declared transpose, never an approximation, and never fires on a path carrying operator parameters (ScalingOp, Advection with a Field coefficient) — test/autodiff.jl guards dκ (skipped on Julia 1.13 until test(ad): re-enable Enzyme gradient tests skipped on Julia 1.13 #108 lands).
• Rule arguments may not mix GC pointers with inline floats (Julia 1.12 calling convention); rule bodies may not allocate. Both are documented in DESIGN §6 and CLAUDE.md.
• A local macOS/aarch64 green run is not evidence about Linux x86_64. Push and read the CI matrix.

—
🤖 Claude refreshed this one: #40 left, #108 joined.
_{claude-opus-5-5[1m] (high) · claude-code 2.1.285 · unreviewed}