Use GetOpenReputerSubmissionWindows query for reputers

[fstecker]

We had effectively disabled polling for reputers because the GetUnfulfilledNonces call returned too many nonces (it included epochs not ready for reputation yet). This reenables polling, but uses the GetOpenReputerSubmissionWindows event instead, which wasn't available in earlier protocol versions. Should result in a slight improvement of reputer reliability.

---

Summary by cubic

Re-enables reputer polling by querying open submission windows and handles query failures without blocking submissions. Previously get_unfulfilled_nonces returned empty and relied only on EventReputerSubmissionWindowOpened; now it returns open window block heights, and explicit/event-driven submissions still run if polling fails.

• Implement get_unfulfilled_nonces via emissions.query.get_open_reputer_submission_windows(...); returns empty when no windows or nonces is None.
• Catch exceptions in _maybe_submit_impl, log a warning on failure, and continue processing the provided nonce.
• Compatibility: nodes must expose GetOpenReputerSubmissionWindows for polling; on older protocol versions, only event-driven submissions will run.

^{Written for commit 5800962. Summary will update on new commits.}

[cubic-dev-ai]

Review completed against the latest diff

<sub>Reply with feedback, questions, or to request a fix.

Fix all with cubic | Re-trigger cubic</sub>

[helder-moreira]

Multi-Agent Review — PR #96

Reviewers: Kimi K3 + GPT-5.6-Sol
Verdict: REQUEST_CHANGES — critical backward-compatibility regression

The new GetOpenReputerSubmissionWindows query and block_height → nonce mapping are correct against the generated v10/v3 protos, the return type matches set[int], and 4/4 tests pass. However, the change breaks the event-driven submission fallback on older nodes.

---

Critical — query failure breaks event-driven submissions (GPT-5.6-Sol)

src/allora_sdk/worker/reputer.py line 123 — The old code returned set() (empty set), so EventReputerSubmissionWindowOpened could still trigger submissions via _maybe_submit_impl (line 685 of worker.py): the event nonce was added to the empty set and submitted. The new code makes an RPC call that throws UNIMPLEMENTED on older nodes. Since _maybe_submit_impl calls get_unfulfilled_nonces() before adding the event nonce, the exception aborts the entire submission — including the event-driven path that was previously working.

The polling loop catches this (except Exception in _polling_worker), so nothing crashes. But the event handler also catches it (except Exception in _handle_submission_window_opened_event), meaning the event nonce is silently dropped. On older nodes, reputer submission is completely disabled — worse than before, when at least events worked.

Fix: Catch the unsupported-method error and return an empty set so the event-driven path remains functional:

try:
    resp = await self.client.emissions.query.get_open_reputer_submission_windows(...)
except Exception:
    return set()  # older nodes: fall back to event-driven only

Warning — reputer_nonce may be None (Kimi K3)

src/allora_sdk/worker/reputer.py line 128 — In the generated v3 protos, ReputerRequestNonce.reputer_nonce is Nonce | None (optional message field). If a node returns an entry without reputer_nonce set, the comprehension raises AttributeError. Add a defensive filter:

return {x.reputer_nonce.block_height for x in resp.nonces.nonces if x.reputer_nonce is not None}

Suggestion — clearer error message for unsupported nodes (Kimi K3)

src/allora_sdk/worker/reputer.py line 123 — The PR body acknowledges older nodes won't implement this RPC, but a raw gRPC UNIMPLEMENTED error in the logs is hard to diagnose. A wrapped error message (e.g. "node does not support GetOpenReputerSubmissionWindows; upgrade your node") would help operators.

Suggestion — missing regression test for unsupported-node fallback (GPT-5.6-Sol)

tests/test_reputer_submit.py — Add a test where get_open_reputer_submission_windows raises the older-node unsupported error and verify the fallback returns an empty set (after the fix above). Existing tests cover populated and empty responses correctly.

(Reviewed by Helder's Agent — Kimi K3 + GPT-5.6-Sol)

[cubic-dev-ai]

All reported issues were addressed across 1 file (changes from recent commits).

<sub>Tip: Review your code locally with the cubic CLI to iterate faster.

Fix all with cubic | Re-trigger cubic</sub>