fix(compiler): inspect sibling datatype instances independently

[oxura]

Fixes #1513.

ty_holds used a datatype-name set as a global visited set. After opening Box<Num{}>, it skipped the fields of the sibling Box<Fun{}> and missed the host-minted closure. Track active datatype expansions instead: remove the name after scanning its constructor fields. Recursive-name cycle protection remains, but sibling value-indexed instances are inspected independently. No type-key serialization, copied visited sets, checker changes, runtime workaround, or special-case Box/Tag handling.

Adds the original two-Box closure-bang program as a permanent regression for the existing GPU-enabled test gate. The helper's source documentation explains the active-path invariant. Protected bend2/bend.ts, hand-written guide/GUIDE.md, release-only CHANGELOG.md and gates/caps are unchanged.

Real behavior proof

Linux x86_64, Bun CLI 1.3.13, Node v22.23.1, clang 22.1.8. Based on current main 381dc877.

• Actual CLI and native C (--gpu off --threads 4) run the reported program and print 42. CPU execution is a control: it already succeeds on main and does not reproduce the original Metal failure.
• Build the same program to C with current main and the fixed compiler. Run actual CUDA-branch preprocessing, clang -E -P -x c -D__CUDACC_RTC__=1 program.c, without emulating a GPU or changing its guards. The host-minted closure is FID 4 in these artifacts. Before: the device evaluator has no case 4 and retains the missing-function error default. After: case 4 is included in the real device dispatch. This proves the changed compiler path, not a GPU runtime pass.
• 17 existing/new closure, recursive indexed/pair/list, and current plain-number Nat seam CLI executions match their expected outputs.
• bun gates/repo.ts: 54/54.

An earlier real JS BigInt-seam reproduction was removed from this patch after upstream #1506 changed Nat host values to numbers and removed the old converter. It is not repinned to an incidental new default.

Validation limits

Metal/CUDA hardware and the 48-mini distributed gate were unavailable. The permanent original regression must run on the existing upstream GPU-enabled gate; no actual device success or performance is claimed here. The preprocessing comparison is a throwaway artifact check, not a permanent source-text test.

[Lorenzobattistela]

Thanks for this.
I am chasing a different approach on #1542 , so I'll close this one in favor of that.