fix(ffi): constructors return JSON handle/error envelope, replacing aimux_last_error (issue #17)

[cunninghamcard-bit]

Summary

Follow-up simplification of #19: constructors return the JSON envelope the rest of the ABI already uses, instead of a bare u64 plus the aimux_last_error() side channel:

{"handle":<u64>}                                     on success
{"error":"...","error_type":"...","status_code":...} on failure

Success and failure arrive in the same return value of the same call, so there is no second call to correlate — the entire last_error mechanism (TLS slot, begin_constructor/record_error, the aimux_last_error symbol and its thread-affinity contract) is removed. No binding needs thread pinning (LockOSThread gone in Go; nothing for Kotlin coroutines / virtual threads to get wrong when they adopt it). The convention matches aimux_generate_text / aimux_embed, so bindings reuse their existing envelope parsers.

Net −168 lines vs master (+1426/−1594) while extending coverage from 3 bindings to all 6.

Error envelopes are byte-identical to #19's (InvalidArgument/Json/UnknownProvider, serde detail for prompt_json/config_json) — the only change is the retrieval mechanism.

Cost: breaking C ABI change — 40+ constructor signatures change from uint64_t to char * (freed with aimux_free_string). 0.2.x is the cheapest moment this will ever be.

Verification

• cargo test -p aimux-ffi --release: 14 passed — fix(ffi): preserve detailed errors across the C ABI (issue #17) #19's error_detail regression cases kept (reading the error from the constructor result); the TLS-semantics cases have no counterpart and are replaced with envelope-semantics tests (per-call error independence, 8-thread concurrent constructors, no pinning)
• Go: go vet + go test ./... with a real cgo link; C/C++ examples compile; swift build passes
• E2E against a live OpenAI-compatible endpoint through the Go and C bindings: unknown-provider detail (full provider list), invalid prompt_json serde detail, real generate, real streaming — all pass
• Java/Kotlin/Flutter: static-only changes following existing patterns; CI to verify

🤖 Generated with Claude Code

[eric8810]

is this in draft or complete?

[cunninghamcard-bit]

是完整的，不是 draft——6 个绑定（Go/Java/Kotlin/Swift/Flutter/C）全部迁移完，#19 的 error_detail 回归测试也移植过来了（14 个通过，信封内容对齐：InvalidArgument/Json/UnknownProvider）。本地验证：Rust 测试、Go 真实 cgo 链接、C/C++ example 编译、swift build 都通过，另外用真实的 OpenAI 兼容 endpoint 走 Go 和 C 绑定跑了端到端（未知 provider 详情、invalid prompt_json 的 serde 详情、真实生成、流式）。Java/Kotlin/Flutter 是纯静态改动，以 CI 为准。

不过 #19 合入后这个 PR 现在有冲突，我 rebase 一下。rebase 之后这个 PR 的性质就变成：在 #19 的基础上进一步简化——删掉 aimux_last_error() 的线程局部机制，构造函数直接返回 JSON 信封，成功和失败在同一次调用的返回值里，没有第二次调用需要关联，各绑定也就不需要处理线程亲和（Go 不需要 LockOSThread，Kotlin 协程/虚拟线程接入时也没有坑）。#19 的测试会改写成信封版保留。rebase 完我在这里说一声。

[cunninghamcard-bit]

rebase 完成。现在这个 PR 基于合入 #19 之后的 master，内容就是把 last_error 侧通道整套换成信封返回：相对 master 净删 168 行（+1426/−1594），同时把绑定覆盖从 3 个扩到 6 个。错误信封内容和 #19 完全一致，测试 14 个全过（#19 的回归用例保留，TLS 语义的几个换成了信封语义的等价测试）。rebase 后重新跑了 Rust 测试、Go 真实链接测试、C/C++ 编译、swift build，以及 Go/C 两个绑定的真实 endpoint 端到端，全部通过。

[eric8810]

ci failed

[cunninghamcard-bit]

CI 绿了——之前挂的是新增测试文件的 rustfmt 格式问题，已修（折进了原 commit，历史不变）。现在全部 job 通过，包括我本地没法验证的 Java（JDK 8/11/17/21）。