Phase 21.7 normalization: optimization pre-work + bench harness expansion

- Add opt-in optimizations (defaults OFF)
  - Ret purity verifier: NYASH_VERIFY_RET_PURITY=1
  - strlen FAST enhancement for const handles
  - FAST_INT gate for same-BB SSA optimization
  - length cache for string literals in llvmlite
- Expand bench harness (tools/perf/microbench.sh)
  - Add branch/call/stringchain/arraymap/chip8/kilo cases
  - Auto-calculate ratio vs C reference
  - Document in benchmarks/README.md
- Compiler health improvements
  - Unify PHI insertion to insert_phi_at_head()
  - Add NYASH_LLVM_SKIP_BUILD=1 for build reuse
- Runtime & safety enhancements
  - Clarify Rust/Hako ownership boundaries
  - Strengthen receiver localization (LocalSSA/pin/after-PHIs)
  - Stop excessive PluginInvoke→BoxCall rewrites
- Update CURRENT_TASK.md, docs, and canaries

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

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

src/llvm_py/instructions/newbox.py

@@ -87,6 +87,12 @@ def lower_newbox(
                 resolver.newbox_string_args = {}
             # Map the resulting box handle to the string argument
             resolver.newbox_string_args[dst_vid] = args[0]
+            # Hint downstream passes that this dst is string-ish
+            if hasattr(resolver, 'mark_string'):
+                try:
+                    resolver.mark_string(int(dst_vid))
+                except Exception:
+                    resolver.mark_string(dst_vid)
         except Exception:
             pass  # Silently ignore failures