fix(codegen): better union-arm selection for object-literal return

[cs01]

User impact

Functions returning a union of interfaces (Hit | Miss, etc.) now pick the correct arm when the returned object literal uses a subset of the arm's fields. Previously the heuristic matched on field-count equality, so any literal with fewer fields than the intended arm silently fell through to the first arm — producing wrong GEP indices and occasional crashes in downstream code.

Change

src/codegen/llvm-generator.ts (return-statement path): when the declared return type is a union and no discriminant match is found, search for the smallest arm whose declared field names contain every key of the returned literal. Exact-size superset wins outright. Fallback to prior first-non-null arm when no superset is found.

Part of the struct-layout-unification effort (plan phase P2). Self-hosting (stage 0/1/2) + full test suite green.

Test plan

• added fixture: tests/fixtures/interfaces/union-return-subset-fields.ts — exercises subset-matching between Big (3 fields) and Small (1 field) arms
• npm run verify — tests + stage 1 + stage 2 all green

[github-actions]

Benchmark Results (Linux x86-64)

Benchmark	C	ChadScript	Go	Node	Place
Binary Trees	1.085s	0.968s	2.066s	0.951s	🥈
Cold Start	0.6ms	0.7ms	0.9ms	21.4ms	🥈
Fibonacci	0.705s	0.706s	1.343s	2.608s	🥈
File I/O	0.073s	0.078s	0.073s	0.143s	🥉
JSON Parse/Stringify	0.003s	0.004s	0.012s	0.011s	🥈
Matrix Multiply	0.390s	0.914s	0.412s	0.345s	#4
Monte Carlo Pi	0.341s	0.348s	0.355s	2.011s	🥈
N-Body Simulation	1.387s	1.746s	1.767s	1.845s	🥈
Quicksort	0.190s	0.218s	0.188s	0.232s	🥉
SQLite	0.250s	0.264s	—	0.314s	🥈
Sieve of Eratosthenes	0.011s	0.021s	0.016s	0.032s	🥉
String Manipulation	0.006s	0.015s	0.012s	0.031s	🥉

CLI Tool Benchmarks

Benchmark	ChadScript	grep	node	xxd	Place
Hex Dump	0.345s	—	0.772s	0.109s	🥈
Recursive Grep	0.016s	0.008s	0.078s	—	🥈