fix(py-sdk): reject slot accesses that do not fit in a slot 🐛

[kriss39]

What

Slot.read / Slot.write in genlayer-py-std pass the byte offset straight to _genlayer_wasi.storage_read / storage_write, whose PyArg_ParseTuple format "y*Iw*" / "y*Iy*" uses the I converter — documented as "without overflow checking". An offset of 2**32 + 5 reaches the executor as 5, -1 as 0xffffffff, 2**40 + 7 as 7. That defeats the executor's slot_access_fits guard (wasi/genlayer_sdk/mod.rs), which exists precisely so an out-of-slot access becomes Errno::Inval: the guest never sends the out-of-range offset, it silently reads or overwrites another location in the same slot.

Offsets are plain Python ints built by pointer arithmetic (VLA.__getitem__: self._off + 4 + idx * size; Slot.read/write are public), so e.g. vla.set_length(2**31); vla[2**30] on a 32-byte element aliases offset (4 + 2**35) mod 2**32 == 4. Slot.indirect is already protected by off.to_bytes(4, 'little'); read/write were not.

Change (executor line v0.3, runners/genlayer-py-std)

• storage/core.py: add SLOT_SIZE = 1 << 32 and check_slot_access(off, len) mirroring the executor's slot_access_fits (off >= 0, len >= 0, off + len <= SLOT_SIZE), called from Slot.read and Slot.write; raises OverflowError, consistent with what indirect already raises.
• tests/test_storage_core.py: in-range boundaries accepted; SLOT_SIZE, SLOT_SIZE - 3 + 4, 2**40 + 7, negative offset and negative length rejected on both read and write (via InmemManager, which would otherwise try to allocate the wrapped range).
• runners/support/versions/current.nix: refreshed genlayer-std, py-genlayer, py-genlayer-multi hashes.

Executor commit: kriss39/genvm-executor@90d4ef1 on pr/v0.3/fix/slot-access-bounds (rebased onto the current pinned 7e0936a).

The root cause in the C shim (runners/cpython/modules/_genlayer_wasi/genlayer.c, "I" → parse as object + PyLong_AsUnsignedLongLong + > UINT32_MAX check) is worth fixing too, but it needs a cpython runner rebuild, which I cannot produce here; the Python-side guard covers the only caller. Happy to add the C change if you prefer it in this PR and can refresh the cpython hash.

Verification

• runners/genlayer-py-std: pytest tests --ignore=tests/embeddings → 978 passed, 1 failed (test_render_rejects_malformed_image, PIL not installed in my venv — unrelated).
• PyArg_ParseTuple('I') masking behaviour confirmed against CPython (PyLong_AsUnsignedLongMask).
• ruff format / ruff check clean on the changed files.
• Runner hashes: I cannot run hash-updater.py here (no nix; runners are not built on macOS). I reproduced the content-addressing locally (make-zip.py under Python 3.13 over the cleanSource-filtered src/ + runner.json, then the wrapper Depends uids) and confirmed it reproduces the three currently committed hashes bit-for-bit before applying it to the patched tree. Please treat the #runners-all nix build as unrun on my side; if it reports a mismatch I will update from the got: values.
• Not built here: the Rust executor crate (unchanged by this PR).

v0.2.x does not vendor the Python runner sources, so no pr/v0.2/... branch.

Independent of the other two PRs I opened today; whichever current.nix change lands second I will rebase and recompute.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Repository: genlayerlabs/genvm-manager/.coderabbit.yaml

Review profile: CHILL

Plan: Advanced

Run ID: e799e104-3b1e-480a-bc1d-6fe02245ff97

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Autopilot is currently an internal CodeRabbit preview.

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

GenVM PR actions

Tick a box to run it (the box unticks itself when handled). Actions only run while the PR has the ci-safe label.

• Force run full tests
• Provision executor PRs

Commands

• /genvm-run-tests — run full tests once for the current manager snapshot
• /merge — queue the exact manager snapshot through the App-owned E2E merge train

[kp2pml30]

Hello, I do not see much value in this check, because it is checked on the rust level. Moreover, this is a big breaking change, so I will leave it as is for now