#!/usr/bin/env bash
# apply.sh <TensorFold>/src -- apply the prefix-cache patch series in order.
# Every patch refuses unless its anchor matches exactly once, so a partial tree fails loudly.
set -euo pipefail
SRC="${1:?usage: apply.sh /path/to/TensorFold/src}"
HERE="$(cd "$(dirname "$0")" && pwd)"
CUDA="$SRC/tensorfold/families/glm_moe_dsa/cuda"
DRAFT="$SRC/tensorfold/families/glm5_next/cuda"
for f in "$CUDA/runner.py" "$CUDA/engine.py" "$DRAFT/dflash2.py"; do
  [ -f "$f" ] || { echo "missing: $f" >&2; exit 2; }
done
echo "==> installing prefix_cache.py"
cp -f "$HERE/prefix_cache.py" "$CUDA/prefix_cache.py"
P=python3
echo "==> 1/7 prefill offset";        $P "$HERE/patch_01_prefill_offset.py" "$CUDA/runner.py" --apply
echo "==> 2/7 drafter resume";        $P "$HERE/patch_02_drafter_resume.py" "$DRAFT/dflash2.py" "$CUDA/runner.py" --apply
echo "==> 3/7 engine wiring";         $P "$HERE/patch_03_engine_wiring.py" "$CUDA/engine.py" --apply
echo "==> 4/7 chunk alignment";       $P "$HERE/patch_04_chunk_alignment.py" "$CUDA/runner.py" --apply
echo "==> 5/7 fixed chunks";          $P "$HERE/patch_05_fixed_chunks.py" "$CUDA/runner.py" --apply
echo "==> 6/7 finalize + bind";       $P "$HERE/patch_06_finalize.py" "$CUDA/prefix_cache.py" "$CUDA/runner.py" "$CUDA/engine.py" --apply
echo "==> 7/7 canonical end";         $P "$HERE/patch_07_canonical_end.py" "$CUDA/prefix_cache.py" "$CUDA/runner.py" "$CUDA/engine.py" --apply
find "$SRC" -name __pycache__ -type d -exec rm -rf {} + 2>/dev/null || true
echo "==> done. Verify imports:"
echo "    PYTHONPATH=$SRC python3 -c 'import tensorfold.families.glm_moe_dsa.cuda.runner as R; print(R.FIXED_CHUNKS, R.PREFIX_ON)'"