1. **TensorFlow Upgrade to v2.21.0 & Toolchain Adjustments**
   * Upgraded the `@org_tensorflow` dependency from `2.20.0` to `2.21.0`.
   * Reordered repository initialization in `WORKSPACE` so `tf_workspace3` runs before hermetic Python initialization (required by TF 2.21+ to define `@xla` and `@tsl`).
   * Dropped support for Python 3.9 (removed `requirements_lock_3_9.txt`).
   * Consolidated multiple separate `pywrap` dependencies under a unified `//tensorflow_text/core/pybinds:pybinds_library` target.
   * Applied necessary compilation patches to various TensorFlow subsystems (framework, common runtime, grappler optimizers, device set, etc.).
   * Updated `@local_xla` references to `@xla`.

2. **Kernel Test Dependency Cleanup**
   * Removed redundant `_for_test` wrapper libraries in `tensorflow_text/core/kernels/BUILD` (such as `constrained_sequence_op_for_test` and `ngrams_op_for_test`).
   * Updated C++ unit tests to link directly against standard op C++ libraries (e.g., `//tensorflow_text:constrained_sequence_op_cc`), eliminating duplicate global initializer linkage conflicts.

3. **Pip/Configure Script Fixes**
   * Explicitly imported `compile_pip_requirements` from `@rules_python//python:pip.bzl` in `oss_scripts/pip_package/BUILD` to resolve dependency resolution errors during configuration.
---------

Co-authored-by: Antonio Sanchez <cantonios@google.com>

PiperOrigin-RevId: 949606721

PiperOrigin-RevId: 951572341

…ries.

- Regenerate `normalization_data.c.oss` using ICU 77.1 with `filters.json` enabled ("normalization": "include") rather than bundling the entire unfiltered ICU data archive. This reduces the source data file from ~91 MB to ~616 KB (and compiled in-binary `.rodata` from ~32 MB to ~220 KB across dependent shared libraries).
- Add shared library symbol stripping (`strip -x` on macOS, `strip --strip-unneeded` on Linux) in `build_pip_package.py` to eliminate unnecessary DWARF debug sections and local symbol tables from wheels.
- Drops the final `.whl` package size from ~60-95 MB down to ~5-7 MB, resolving PyPI upload limits.

PiperOrigin-RevId: 967966028

Reduce TF-Text wheel size by filtering ICU 77 data

In ICU 77, BUILD.bazel files were added upstream under icu4c/source/stubdata/ and icu4c/source/i18n/. Because custom_http_archive only excluded icu4c/source/common/BUILD.bazel, Bazel treated icu4c/source/stubdata/ as a separate package, preventing glob(["icu4c/source/stubdata/*.cpp"]) in //third_party/icu:BUILD.bzl from matching stubdata.cpp. Consequently, U_ICUDATA_ENTRY_POINT (icudt77_dat) was omitted from the build, resulting in undefined symbol errors upon importing tensorflow_text.

Excluding stubdata/BUILD.bazel and i18n/BUILD.bazel and including stubdata/*.h in hdrs allows stubdata to be cleanly compiled into icuuc, resolving icudt77_dat.

Fixes #1541

Exclude stubdata and i18n BUILD.bazel from ICU 77 repository archive