Inference-first annotation rules: BSK-0001/0002 never fire on inferable types

CLAUDE.md

@@ -11,7 +11,11 @@ Code here must comfortably pass review at a top-tier engineering org. Keep quali
 
 ## Conformance Is the Prime Directive
 
-Target: **100% PEP conformance**, canonical Python **3.12**. Read the [PEP conformance README](https://github.com/python/typing/blob/main/conformance/README.md) carefully. This discipline outranks every other concern in this file.
+Target: **100% conformance with the maintained Python typing specification**.
+Python-version boundaries apply only where the typing specification, an
+accepted PEP, or Python language semantics defines one; Basilisk has no
+canonical Python release. Read the [PEP conformance README](https://github.com/python/typing/blob/main/conformance/README.md)
+carefully. This discipline outranks every other concern in this file.
 
 - **One reproducible path — the real harness.** `python3 conformance/run_conformance.py` clones `python/typing@main` FRESH and runs the suite's OWN unmodified harness (`src/main.py --only-run basilisk`) over the binary in its default config — every PEP rule on, nothing configured ([CHKARCH-CONFORMANCE], [CHKARCH-CONFIGURATION-ONLY]). The score is exactly what a user gets out of the box; never quote a number produced any other way, and never re-introduce a home-grown/vendored scorer.
 - **Precision is the whole game.** A file passes iff the upstream `errors_diff` is empty: emit an error on EVERY `# E` line, satisfy EVERY `# E[tag]` group, and emit NOTHING on a line the suite does not mark. Follow each PEP exactly — no missed required error, no stray diagnostic.

README.md

@@ -42,12 +42,12 @@ And it is the **fastest checker we’ve measured** — median cold full-
 
 | Type checker | Median cold check |
 | --- | --- |
-| ⚡ **Basilisk** | **<!--g:benchBasilisk-->11<!--/g:benchBasilisk--> ms** |
-| zuban | <!--g:benchZuban-->27<!--/g:benchZuban--> ms |
-| ty | <!--g:benchTy-->39<!--/g:benchTy--> ms |
-| Pyrefly | <!--g:benchPyrefly-->148<!--/g:benchPyrefly--> ms |
-| Pyright | <!--g:benchPyright-->582<!--/g:benchPyright--> ms |
-| mypy | <!--g:benchMypy-->610<!--/g:benchMypy--> ms |
+| ⚡ **Basilisk** | **<!--g:benchBasilisk-->10<!--/g:benchBasilisk--> ms** |
+| zuban | <!--g:benchZuban-->28<!--/g:benchZuban--> ms |
+| ty | <!--g:benchTy-->40<!--/g:benchTy--> ms |
+| Pyrefly | <!--g:benchPyrefly-->110<!--/g:benchPyrefly--> ms |
+| Pyright | <!--g:benchPyright-->573<!--/g:benchPyright--> ms |
+| mypy | <!--g:benchMypy-->574<!--/g:benchMypy--> ms |
 
 Median cold full-file check across <!--g:benchCount-->26<!--/g:benchCount--> single-construct typing-spec stress fixtures on an <!--g:benchMachine-->Apple M4 Max<!--/g:benchMachine--> &mdash; lower is better. Basilisk&rsquo;s warm re-check drops to ~<!--g:benchWarm-->4<!--/g:benchWarm--> ms. Every figure is produced by [`hyperfine`](https://github.com/sharkdp/hyperfine) and committed per machine, so nothing here is hand-typed. **Clone the repo, run `make bench` on your own hardware, and send us the CSV &mdash; independent audits are welcome.** [Full benchmarks &amp; methodology &rarr;](https://www.basilisk-python.dev/docs/benchmarks/)
 

README.zh.md

@@ -43,12 +43,12 @@ Basilisk 是**唯一**在官方
 
 | 类型检查器 | 冷启动检查中位数 |
 | --- | --- |
-| ⚡ **Basilisk** | **<!--g:benchBasilisk-->11<!--/g:benchBasilisk--> ms** |
-| zuban | <!--g:benchZuban-->27<!--/g:benchZuban--> ms |
-| ty | <!--g:benchTy-->39<!--/g:benchTy--> ms |
-| Pyrefly | <!--g:benchPyrefly-->148<!--/g:benchPyrefly--> ms |
-| Pyright | <!--g:benchPyright-->582<!--/g:benchPyright--> ms |
-| mypy | <!--g:benchMypy-->610<!--/g:benchMypy--> ms |
+| ⚡ **Basilisk** | **<!--g:benchBasilisk-->10<!--/g:benchBasilisk--> ms** |
+| zuban | <!--g:benchZuban-->28<!--/g:benchZuban--> ms |
+| ty | <!--g:benchTy-->40<!--/g:benchTy--> ms |
+| Pyrefly | <!--g:benchPyrefly-->110<!--/g:benchPyrefly--> ms |
+| Pyright | <!--g:benchPyright-->573<!--/g:benchPyright--> ms |
+| mypy | <!--g:benchMypy-->574<!--/g:benchMypy--> ms |
 
 在 <!--g:benchMachine-->Apple M4 Max<!--/g:benchMachine--> 上，跨 <!--g:benchCount-->26<!--/g:benchCount--> 个单一类型构造压力测试样本的冷启动全文件检查中位数 —— 数值越低越好。Basilisk 的热重检查可降至约 <!--g:benchWarm-->4<!--/g:benchWarm--> ms。每个数字均由 [`hyperfine`](https://github.com/sharkdp/hyperfine) 生成并按机器提交，没有一个是手写的。**克隆仓库，在你自己的硬件上运行 `make bench`，并把 CSV 发给我们 —— 欢迎社区独立复核。** [完整基准测试与方法论（英文）&rarr;](https://www.basilisk-python.dev/docs/benchmarks/)