feat(phase21.5): Stage-B parser loop fix + delegate path stabilization

## 修正内容

### 1. Stage-B パーサー修正（偶然の回避）
- **ファイル**:
  - `lang/src/compiler/parser/expr/parser_expr_box.hako`
  - `lang/src/compiler/parser/stmt/parser_control_box.hako`
- **問題**: ネストループで gpos が正しく進まず、loop の cond/body が壊れる
- **回避策**: new 式のメソッドチェーン処理追加で別ループを導入
- **結果**: MIR 生成が変わって VM gpos バグを回避

### 2. delegate パス動作確認
- **テスト**: `/tmp/loop_min.hako` → rc=10 ✅
- **MIR構造**: 正しい PHI/compare/binop を生成
- **チェーン**: hakorune parser → Rust delegate → LLVM EXE 完動

### 3. ドキュメント追加
- `docs/development/analysis/` - delegate 分析
- `docs/development/guides/` - ループテストガイド
- `docs/development/testing/` - Stage-B 検証報告

### 4. カナリーテスト追加
- `tools/smokes/v2/profiles/quick/core/phase2100/` 配下に複数追加
- emit_boxcall_length_canary_vm.sh
- stageb_parser_loop_json_canary_vm.sh
- 他

### 受け入れ基準
- ✅ delegate パス: rc=10 返す
- ✅ FORCE パス: rc=10 返す（既存）
- ✅ MIR 構造: 正しい PHI incoming と compare
- ✅ 既定挙動: 不変（dev トグルのみ）

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

docs/development/roadmap/phases/phase-21.5-optimization/CHECKLIST.md

@@ -0,0 +1,6 @@
+- [ ] Baselines captured (loop/strlen/box), median/avg recorded
+- [ ] VM stats correlated (inst/compare/branch)
+- [ ] Structural optimizations applied (const/temps dedupe)
+- [ ] Re-measure after each change (no behavior change)
+- [ ] EXE canaries remain PASS (loop/print/strlen FAST)
+- [ ] README.ja/quick guides updated when toggles affect usage

docs/development/roadmap/phases/phase-21.5-optimization/README.md

@@ -0,0 +1,31 @@
+# Phase 21.5 — Optimization (ny-llvm crate line)
+
+Scope
+- Optimize hot paths for the crate (ny-llvmc) line using Hakorune scripts only.
+- Preserve default behavior; all risky changes behind dev toggles.
+- Measure EXE runtime (build once, run many) to avoid toolchain overhead noise.
+
+Targets (initial)
+- loop: integer accumulations (no I/O)
+- strlen: FAST=1 path (pointer → nyrt_string_length)
+- box: construct/destroy minimal boxes (String/Integer)
+
+Methodology
+- Build once via ny-llvmc; time execution only (`--exe` mode).
+- Runs: 3–5; report median and average (target ≥ 100ms per run).
+- Observe NYASH_VM_STATS=1 (inst/compare/branch) where relevant to correlate structure and runtime.
+
+Commands (examples)
+- tools/perf/phase215/bench_loop.sh --runs 5
+- tools/perf/phase215/bench_strlen.sh --runs 5 --fast 1
+- tools/perf/phase215/run_all.sh --runs 5 --timeout 120
+
+Dev Toggles (keep OFF by default)
+- NYASH_LLVM_FAST=1 (strlen FAST)
+- HAKO_MIR_BUILDER_JSONFRAG_NORMALIZE=1 (normalize)
+- HAKO_MIR_BUILDER_NORMALIZE_TAG=1 (tag, test-only)
+
+Exit Criteria
+- Representative microbenches stable (≤ 5% variance) and ≥ 80% of C baselines.
+- No regression in EXE canaries (loop/print/strlen FAST) and VM parity canaries.
+