perf(vm): bypass runtime coercion for primitive operands

[Eric-Song-Nop]

Stacked on #12. Base: perf/static-property-atoms; head: perf/primitive-coercion-fast-path, both in pocket-stack/quickjs-oxide. Review and merge in dependency order.

Migrated from Eric-Song-Nop#3 with identical commits and preserved measurement/validation evidence.

Change / 修改

Arithmetic and comparison instructions used to send primitive operands through VmHost into Runtime::to_primitive before receiving the same value back. Add one inline VM ToPrimitive entry point: return primitives unchanged and delegate objects, with the original hint, to the existing host. Numeric preparation and unary/add/comparison instructions share that entry point; arithmetic, result tagging, string/BigInt handling and object conversion remain single implementations.

数值与比较指令此前把原始值送入 VmHost/Runtime 后再原样返回。现在通过一个共享 VM ToPrimitive 入口直接返回原始值，对象及原 hint 交给现有 host。数值准备与一元/加法/比较复用入口，计算、结果标签、字符串/BigInt 与对象转换规则保持单一实现。

Related issue / 关联 issue: #8

Validation / 验证

• 3,002 Rust workspace tests passed; 0 failed; 1 existing ignored. Tests exercise primitive host bypass, representation/NaN-bit preservation, object hints/order/abrupt completion and mixed numeric semantics. / 3002 项通过，0 失败，1 项原有忽略。
• Formatting, source layout, Rust-only and strict production Clippy passed in profiling/default/test262-host configurations. / 格式、布局、Rust-only 和三种配置的生产目标 strict Clippy 通过。
• Full Test262 ran on both revisions: 102,037 outcome rows identical, zero new failures. Only the first source-fingerprint header differs; both frozen full-file receipt gates still exit 1, and no baseline was updated. / 前后完整 Test262 结果逐行一致；冻结文件校验仍因源码指纹头变化失败，没有修改基线。
• Binary-object production boundary passed; all 690 isolation canaries rejected. / 字节码生产边界检查通过，690 个隔离 canary 全部按预期拒绝。

Measurements / 测量

Same PocketLab machine and release flags, runtime diagnostics inactive, unchanged workloads. All 80 runs passed output checks: five alternating pairs for each initial workload plus a separate ten-pair confirmation for each Richards workload, using the exact same binaries. All initial and confirmation samples are retained.

同机同 release 配置、运行时诊断关闭、工作负载未改动。首轮每负载五轮交替对照，随后对两种 Richards 分别独立复测十轮，使用哈希相同的二进制。80 次输出全部校验通过，保留两轮全部样本。

Workload / 负载	Before median ms	After median ms	Time reduction / 耗时下降
Integer loop ×10 million / 整数循环一千万次	1967.352	1547.222	21.36%
Property reads ×20 million / 属性读取负载两千万次	7041.108	6522.871	7.36%
Fixed Richards, initial / 固定 Richards 首轮	1496.602	1506.159	-0.64%
Original Richards, initial / 原始 Richards 首轮	6243.921	6295.843	-0.83%
Fixed Richards, confirmation / 固定 Richards 复测	1498.446	1494.523	+0.26%
Original Richards, confirmation / 原始 Richards 复测	6224.281	6230.859	-0.11%

Richards ranges overlap in both rounds; no stable Richards gain is established, and a small regression remains possible. Property-read timing includes numeric loop/accumulation work. CPU frequency/system isolation were not controlled; this is not a whole-engine speedup claim. No new CPU sampling or phase/memory probe was performed.

Richards 两轮范围均重叠，没有建立稳定收益，小幅回退仍有可能。属性读取负载包含数值循环和累加；未锁频或完全隔离系统，不把单个负载收益推广到整个引擎。本轮未重新采集 CPU 栈或阶段/内存探针。

双语报告 / Report · 全部样本及验证记录 / Samples and receipts