[-16 lines] Reuse compiler paths for array access and branch blocks

[PedroVIOliv]

Two compiler paths repeat logic already shared elsewhere. arr_op repeats array ownership/alias setup and the old-cell read and tuple return for get/swap. emit_chain implements its own indentation and closing braces instead of using block.

Share the existing-array setup and cell read/result path, preserving get's retain, swap's move and set's sink. Emit each conditional arm through block, retaining the single-arm shortcut. Generated else clauses move to the next line; conditions and branch bodies stay the same.

Only bend2/comp.ts changes: 20 insertions, 36 deletions (net -16 lines, -340 bytes). No new helpers, feature changes, comment removal or size-cap changes. Measured with the gate's default ttok invocation: 64,941 tokens upstream, 64,894 after the array change, and 64,854 after both changes (-87 total). The compiler is 146 tokens below its 65,000-token cap.

Validation against upstream e52cda47:

• Compared compiler exit status, stdout, stderr and emitted C/JS for 779 Base-importing test/benchmark programs with a main and no expected compile error. All matched after normalizing only the closing-brace/else line break: 639 successful programs and 140 identical compiler rejections. Of the successful programs, 564 have formatting differences; 75 remain byte-identical.
• All 16 runtime benchmarks produced identical optimized CPU LLVM IR with Apple clang 17 at -O3, excluding only the source path and module identifier. Runtime benchmark timings were not measured.
• All 139 runnable tests selected from tests/compile and array-named tests passed their existing #| expectations on JS and C compiled with Apple clang 17 at -O3, using one CPU thread. Expected nonzero exit statuses were checked as well.
• Metal: all 19 noninteractive GPU-bearing tests from the successfully emitted corpus passed on both upstream and this branch (38 runs) on an Apple M3, with --gpu on --threads 1. Every build created a Metal pipeline archive. This includes array atomics/forks, closure reachability, fork leaf continuations, GPU trig and stencil3d. The remaining GPU-bearing test, tests/gfx/clicks.bend, requires interactive window events and was excluded.
• git diff --check passed.
• The full repository size gate passed: PASS: 46 / 46, using ttok 0.3 installed in an isolated temporary environment.

The cluster test gate could not run because this host cannot resolve cluster. Cluster test/performance gates remain for maintainer infrastructure; CUDA execution and GPU performance measurements were not tested locally.

[VictorTaelin]

The first commit (the shared cell-access path in arr_op) is merged into 2.0.25 under your authorship, thank you: −47 ttok, byte-identical emission. The second (emitting conditional arms through block) moves every else in the emitted C and JS to its own line; that is a formatting call for Victor, so I left it out for now. If he wants it, we will take it from here.

Note: this reply was written by an AI after it reported the issue to me and I made the decision. If anything here is wrong, reply and I will review it myself.