v0.8.2 — CLI cleanup: push/pull + drop reward/init

End-to-end-validated via a live HF Hub round-trip on pallets/click (no E2B / cloud sandbox).

CLI changes

Action	Command
Add	repo2rlenv push <local-dir> <hf://owner/dataset> — supports --private, --message
Add	repo2rlenv pull <hf://owner/dataset> [<local-dir>] — supports --task, --force
Remove	repo2rlenv reward (Python function repo2rlenv.reward.calculate_diff_similarity_reward stays for training-loop users)
Remove	repo2rlenv init (replaced by README example)
Soft-deprecate	--out hf://... magic inside generate — emits a warning; to be removed in v0.9

Resulting CLI (5 commands, every one load-bearing):

repo2rlenv generate    # emit Harbor task spec → local dir
repo2rlenv validate    # check spec conformance
repo2rlenv bootstrap   # build a Docker image only (no synthesis)
repo2rlenv push        # NEW: local dir → HF Hub
repo2rlenv pull        # NEW: HF Hub → local dir

Harness fix surfaced + landed in this release

The bootstrap agent sometimes saved test_cmds with a trailing | head -50 pipe; targeted_test_cmds_for_pr then appended PR test files after the pipe → broken shell. normalize_test_cmds_for_runtime now strips shell pipes / redirects (| head, | tail, 2>&1, > /dev/null) before per-runner normalization.

Live round-trip evidence

Step	Result
generate	✅ 4 tasks emitted on pallets/click
validate (local)	✅ 4/4
push hf://AdithyaSK/click-r2e-v082	✅ commit 82de5c57
pull hf://AdithyaSK/click-r2e-v082	✅ 4 tasks materialized
validate (pulled copy)	✅ 4/4
harbor run --agent oracle on pulled copy	✅ Mean = 1.000 in 5s

Test plan summary

• 464/464 unit tests pass (was 442 on v0.8.1; +22 new — 19 push/pull tests + 3 normalize regression tests)
• ruff check + format check clean
• Full GitHub Actions CI green (lint + py3.12/3.13/3.14 + sdist+wheel build)

Closes #16.