fix(windows): align CRT to dynamic to resolve ONNX __imp_ link errors (#7)

The x86_64-pc-windows-msvc release link failed with 75 LNK2019/LNK2001
errors on __imp_* C-runtime symbols. Root cause is a CRT linkage conflict
(pykeio/ort #329): ONNX Runtime's prebuilt lib (ort download-binaries)
links the CRT dynamically (/MD) and references __imp_* import stubs, but
two C build-deps hardcode .static_crt(true), dragging libcmt/libucrt into
the link and leaving ONNX's __imp_* unresolved:

  - esaxx-rs    via tokenizers' default `esaxx_fast` feature (training-only)
  - libsqlite3-sys (bundled) via burn's default `dataset` feature

Both are pulled by features WhisperForge never uses (no .encode(), no
tokenizer training, no burn datasets). Fixes:

- tokenizers: default-features=false + ["onig"] — drops esaxx_fast, so
  esaxx-rs no longer compiles its static-CRT C++. `onig` keeps the same
  regex backend as the default, so load/decode behavior is unchanged.
- .cargo/config.toml: /NODEFAULTLIB:libcmt+libucrt and force msvcrt+ucrt,
  neutralizing libsqlite3-sys's static CRT and returning the link to
  ONNX's standard all-dynamic configuration. The prior
  legacy_stdio_definitions.lib flag targeted the wrong symbols.

Also add a `windows-link-check` CI job that links the MSVC target with the
dist feature set on every PR — the v0.5.3 failure surfaced only at tag time
because PR runs only do `dist plan`, not a real build.

Author: bevsxyz

.cargo/config.toml

@@ -8,9 +8,19 @@
 # Release builds (dist) deliberately use the stock toolchain on ubuntu-22.04 for
 # broad glibc compatibility.
 
-# ORT pre-built static library (ort 2.0.0-rc.10, download-binaries) references
-# old-MSVCRT symbols (__dstbias, __timezone, __tzname) that live only in
-# legacy_stdio_definitions.lib — Microsoft's backward-compat shim for pre-VS2015
-# CRT references, absent from ucrt.lib/msvcrt.lib/vcruntime.lib.
+# ONNX Runtime (ort 2.0.0-rc.10, download-binaries) links the CRT *dynamically*
+# (/MD) and references __imp_* import stubs. `libsqlite3-sys` (bundled sqlite, via
+# burn's default dataset feature) compiles its C with the *static* CRT (/MT), which
+# drags libcmt/libucrt into the link and leaves ONNX's __imp_* unresolved (pykeio/ort
+# #329). Force the whole link onto the dynamic CRT so ucrt.lib/msvcrt.lib supply the
+# __imp_* stubs. The static-CRT C code (sqlite datasets) is never called at runtime,
+# so CRT-mixing has no runtime effect — only the link must be consistent.
+# NOTE: do not set the RUSTFLAGS env var on the Windows build (it overrides these
+# target rustflags); see ci.yml windows-link-check.
 [target.x86_64-pc-windows-msvc]
-rustflags = ["-C", "link-arg=legacy_stdio_definitions.lib"]
+rustflags = [
+    "-C", "link-arg=/NODEFAULTLIB:libcmt.lib",
+    "-C", "link-arg=/NODEFAULTLIB:libucrt.lib",
+    "-C", "link-arg=/DEFAULTLIB:msvcrt.lib",
+    "-C", "link-arg=/DEFAULTLIB:ucrt.lib",
+]

.github/workflows/ci.yml

@@ -53,3 +53,34 @@ jobs:
 
       - name: Test
         run: cargo test --release --locked -p whisperforge-core -p whisperforge
+
+  windows-link-check:
+    name: Windows MSVC link check
+    runs-on: windows-2022
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Install Rust toolchain
+        uses: dtolnay/rust-toolchain@stable
+        with:
+          targets: x86_64-pc-windows-msvc
+
+      - name: Cache Cargo registry and build
+        uses: actions/cache@v4
+        with:
+          path: |
+            ~/.cargo/registry/index/
+            ~/.cargo/registry/cache/
+            ~/.cargo/git/db/
+            target/
+          key: ${{ runner.os }}-cargo-${{ hashFiles('**/Cargo.toml', 'Cargo.lock') }}
+          restore-keys: |
+            ${{ runner.os }}-cargo-
+
+      # NOTE: deliberately do NOT set RUSTFLAGS here — it would override the
+      # .cargo/config.toml [target.x86_64-pc-windows-msvc] rustflags this job exists
+      # to validate. (The `check` job sets RUSTFLAGS=mold, which is why this is separate.)
+      - name: Build wforge (dist profile, gpu) — link only
+        run: cargo build --profile dist --features gpu -p whisperforge --target x86_64-pc-windows-msvc --locked

Cargo.lock

@@ -58,8 +58,13 @@ rustfft = "6.2"
 cpal = "0.15"
 ringbuf = "0.4"
 
-# Text processing
-tokenizers = "0.23"
+# Text processing.
+# default-features disabled to drop the `esaxx_fast` feature, which pulls esaxx-rs —
+# a C++ build-dep that hardcodes `.static_crt(true)`, forcing the static CRT (/MT) and
+# breaking the Windows MSVC link against ONNX Runtime's dynamic-CRT (/MD) prebuilt lib
+# (see .cargo/config.toml and pykeio/ort #329). esaxx_fast is training-only; we only
+# load + decode. `onig` keeps the same regex backend as the default → zero behavior change.
+tokenizers = { version = "0.23", default-features = false, features = ["onig"] }
 
 # Serialization and CLI
 serde = { version = "1.0", features = ["derive"] }