[evaleev]

Follow-ups to the reviews of #399 (the nonblocking items that were still open when it merged).

Engine

• compute1()/compute2() check the shells against the Engine's max_l in all builds and throw Engine::lmax_exceeded. Before, this was only a Debug assert, so a Release build overran the stack/scratch sized for max_l; this is also the case on master before Relativistic 2e 3- and 4-center integrals #399.
• lmax_exceeded::lmax_limit() is the max AM wherever it is thrown; initialize() used to report max AM + 1. what() now names the task and both values.
• rkb_operators docs: the "folded" paragraph applies to the operators with one folded electron; opop_coulomb_opop already contains the −1 of i².
• New unit tests:
  • the per-center check of compute2() (live whenever LIBINT2_ERI3_MAX_AM > LIBINT2_MAX_AM, as in CI);
  • the max_l check;
  • the limit reported by initialize();
  • the core evaluator rebuilt by set(BraKet).
• set(BraKet)/set(Operator) resize the primitive data for the rank of the new braket. Before, initialize() resized it only when given max_nprim, so e.g. switching an Engine from 3 to 4 centers overran the buffer (found by the new set(BraKet) test under ASan). max_nprim() returns the value the Engine was sized for; it used to be derived from the size of the last computed shell pair.
• compute2() evaluates the total derivative order (der + the operator's intrinsic order) at compile time instead of calling intrinsic_deriv_order() per primitive.

hartree-fock++ / CI

• The validator accepted "2-body forces/hessian skipped" unconditionally. The features file now has the ERI max AM for each derivative order (LIBINT_ERI_MAX_AM_DERIVS), and the validator accepts a skip only if the kernels cannot handle the d shells of the basis.
• The CI lane with LIBINT2_MAX_AM="2;2;1" generated 2nd-derivative ERIs up to p only. hartree-fock++ therefore skipped its Hessian, which was tested in no lane. The lane now generates them up to d (LIBINT2_ERI_MAX_AM="2;2;2").
• api_basic_compile_test: the 2-center derivative checks also need the 2eri<d> kernels.
• Fortran ERI test: the difference is measured relative to the largest element of the shell set. With 2nd derivatives over d shells, 2 of ~1400 elements differed by 5e-12 relative to themselves, 3e-15 in absolute terms; this is roundoff from differently ordered prim-data arithmetic.

Compiler

• RKB fold relations: the symbolic sum leads with a +1 term, so the cross × cross components no longer multiply by −1.0. That multiply was also missing from the flop count.
• The component-index range checks of the RKB/σ·p descriptors now run in all builds and check both bounds.
• PairSwap::coupled ⇒ p1p2_swappable is a static_assert instead of asserts in the shell-set predicates.
• SMultipole_1_1_int was outside the LIBINT_SUPPORT_ONEBODYINTS guard.
• Stale comments, unused includes, a leftover flush.
• RecurrenceRelation::is_alias(): an RR whose target is identical to its only child. handle_trivial_nodes() makes such a vertex refer to its child (when the child has no other parent and is not a target), and no code is generated or included for the RR. CR_11_Coulombσp_11 is one, so the coulomb_op kernels accumulate the derivative integrals directly instead of copying them first. In the configuration below, 19 3coulomb_op kernels change, CR_Coulombop_{0,1,2} are no longer generated, and every other generated file is unchanged.
• Compiler core test: new DerivGaussV2 test (children of direct differentiation and of translational invariance, code shared by spectators but not by shells or directions, code-generating builds of both paths, CR_11_Coulombσp_11 is an alias). The test driver now exits nonzero if a test throws; it used to return 0 regardless.

Verified locally (macOS, Apple clang, gfortran 14) with the configuration of the modified CI lane (MAX_AM=2;2;1, ERI_MAX_AM=2;2;2, ENABLE_ERI=2, ERI3_MAX_AM=3;2, RKB on, RKB_ERI3_MAX_AM=3, ERI2 on, Fortran, Release):

• check-libint2compiler passes 2/2 and check passes 17/17: 47 unit test cases, 134,570,582 assertions.
• hartree-fock++ validates the 2-body Hessian; a faked "skipped" line is rejected.
• The per-center section of the new lmax test is live in this configuration (3-center limit 3 > MAX_AM 2).
• All unit tests also pass under AddressSanitizer (Debug headers against the generated library).
• The full library was regenerated and rebuilt after the alias change, with stale generated files removed; check-libint2compiler and check pass again.

This PR was agent-generated and human-reviewed.