Autoharness: skip fns whose const {} blocks depend on a substituted type

[acearyanarun]

#4820 stopped a const {} precondition on a const parameter from aborting an autoharness run. As #4826 describes, the same unrecoverable E0080 was still reachable when the substituted type decides the block: either a const parameter reaching it through a type argument (inner::<[u8; N]>()) or a block guarded on a type parameter alone (const { [..][size_of::<T>()] }). Both abort the entire run.

The change

The reachability walk now carries (DefId, GenericArgsRef), expressed in terms of the candidate's own parameters, instead of bare DefIds. It follows calls, reified fn items, closures and coroutines whose instantiated arguments mention a type or const parameter. Instantiations with concrete arguments are still not followed. A const {} block (AnonConst with NonTypeSystemInline) counts as a possible precondition when its instantiated arguments mention a type or const parameter.

One deliberate departure from the issue's suggested direction: other unevaluated constants still count only when they mention a const parameter. Generic associated constants like <T as SizedTypeProperties>::IS_ZST are reached by nearly every function over slices and collections. Treating them as preconditions would skip all of those (slice_len in the test pins this boundary).

Carrying instantiated arguments is mostly about precision: a caller that reaches a guarded callee only through a concrete type is no longer skipped (concrete_second_type pins this).

Std skip-count impact (requested in #4826)

kani autoharness --std --list over the nightly-2026-09-25 library, on aarch64-darwin, excluding core_arch:: (on main it crashes on SVE tuples; see #4964):

• Generated: 20,868 → 20,790 (−78, 0 gained). Every lost function reaches a type-dependent const {} block, e.g. TypeId::of, needs_drop, type_name, RawVec's const { assert!(..) }, read_unaligned, mem::zeroed.
• The other changed skip rows are relabeling. Functions already skipped for missing Arbitrary or trait bounds now report the const-block reason, because the check runs first, and are listed under their generic rather than instantiated name.
• The 1024-body search limit was never hit (checked with temporary instrumentation).

Testing

• New cases in cargo_autoharness_generics: both shapes from autoharness: E0080 can still abort a run when the substituted *type* decides a const {} block #4826, a type-parameter variant, and three must-still-verify boundaries (concrete type argument, a concrete argument passed through an intermediate call, generic associated constants). The two boundary tests that pin design choices fail under their targeted mutations.
• autoharness_comptime_intrinsic/expected updated for the reworded skip message.
• script-based-pre 90/91 locally. codegen_cache hit an EINTR on echo under parallel load; it passes in isolation and passed in an earlier full run.
• cargo test -p kani-compiler, clippy -D warnings, kani-fmt --check all pass.

Resolves #4826

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 and MIT licenses.

[acearyanarun]

Hey @CYJ904 @Tianshu-Huang @wodex1nhaoIeng @srivatsansamraj, I opened #4981 for Milestone 33 / #4826. Can you guys take a look when you get a chance and sanity check the fix + numbers?