TensorFlow 0.1.3: context-bound constants and default transpose

[rextio]

Summary

• prepare rextio-tensorflow 0.1.3 as an Unreleased candidate and add the 0.1.3 integration branch to CI
• add an exact fail-closed CPU float32 rank-2 tf.transpose(x) lowering using same-wheel TFE Transpose with default [1, 0] permutation only
• add a per-BorrowedContext RAII prepared-constant table for reduction axes, ArgMax axes, and the rank-2 transpose permutation
• explicitly clear cached TFE handles before Python eager-context anchors are released
• preserve TensorHandle payload sharing, TensorFlow 2.21.0 and CPython 3.11 ABI pins, existing operation semantics, and all CUDA non-claims
• keep the CUDA-derived runtime buildable and update new 0.1.3 manual evidence identity without rewriting the historical 0.1.2 record

Safety and evidence boundaries

Prepared handles are never process-global and are never shared across distinct BorrowedContext values, even when separate extracts observe the same raw eager-context pointer. Reuse is limited to repeated operations in one generated native ownership graph. Drop for BorrowedContext drains the prepared table while the strong Python context and capsule anchors are still alive.

The real-Cargo coverage uses a non-square 2x3 tensor, proves 3x2 transpose parity, exercises double transpose and repeated reduction helper paths, checks special values and boundary rejection, and verifies returned tensors remain valid after inputs and Rust locals are released. It does not claim an observed cache-hit rate or speedup.

Verification

• focused candidate suite after review fixes: 389 passed
• focused real-Cargo transpose/RAII test: 1 passed
• earlier full non-GPU suite: 433 passed, 1 skipped, 8 deselected
• earlier full real-Cargo suite: 8 passed
• supervisor recheck: prepared/package 11 passed; transpose claim/lower 11 passed; CUDA identity/static contracts 3 passed
• Ruff: clean
• mypy: clean across 30 source files
• staged diff whitespace check: clean

Deliberate limitations

• no graph or FunctionDef fusion
• no cross-generated-function residency or process-wide cache
• no explicit permutation, conjugate, name, rank-1, non-float32, or CUDA transpose lowering
• no performance or speedup claim
• CUDA remains build-only and non-promoting with support_claim=false and certification_ready=false

This PR targets the 0.1.3 integration branch. It does not tag, release, publish, or merge to main.