Solana plugin: blockhash handling + maintenance audit

[jmoreira-valory]

Problem

The open-aea-ledger-solana plugin has accumulated correctness and maintenance debt that should be addressed in a dedicated refactor. The dep bump in #(this PR) re-enables the plugin for downstream consumers (see valory-xyz/open-autonomy#2479) but intentionally preserves pre-existing behaviour to keep that change minimal and review-able. This issue tracks the items that should be cleaned up next.

Correctness: blockhash handling

The plugin caches a recent blockhash with a wall-clock TTL (10 seconds, both before and after #(this PR) — the upstream solana.blockhash.BlockhashCache being replaced was itself TTL-based). Solana blockhash validity is tied to block height, not time:

• a blockhash expires when current_block_height > last_valid_block_height (~150 blocks, ~60–90s but variable with cluster load)
• context.slot from get_latest_blockhash is the slot the RPC responded at, not an expiration bound
• under load / async delays this yields intermittent Blockhash not found at submit time

Needed:

• Replace the TTL cache with a height-based model:
  • fetch value.blockhash and value.last_valid_block_height from get_latest_blockhash()
  • on get, call get_block_height() and compare against last_valid_block_height — expire if exceeded
  • typed attribute access on the response (no to_json() → json.loads(...) roundtrip)
• OR (simpler, preferred for agents that submit infrequently) remove the cache entirely and fetch per-transaction. SolanaApi.latest_hash already does an unconditional RPC call on every access today, so the cache is mostly ornamental.
• Drop the unused slot parameter from BlockhashCache.set() — it is stored but never read back.

Correctness: other pre-existing issues

• _generate_tx_nonce calls self._api.BlockhashCache.get(), but BlockhashCache is an attribute of SolanaApi, not ._api (a SolanaApiClient — confirmed: hasattr(SolanaApiClient(...), 'BlockhashCache') is False). Every call raises AttributeError, is caught by the broad except Exception, and falls through to a fresh RPC. The cache has been ornamental in practice — removing it entirely in a follow-up changes nothing behaviourally. Correct reference: self.BlockhashCache.get().
• Broad except Exception in _generate_tx_nonce swallows RPC failures, serialization errors, and logic bugs alike. Narrow to LookupError (cache miss) so real failures surface.
• Missing null check on get_latest_blockhash().value — crashes with AttributeError if the RPC returns no value rather than raising.
• File-handle leak in SolanaHelper.load_contract_interface: open(bytecode_path, "rb") without a with block. Add context manager + explicit encoding.
• JSON roundtrip in _get_latest_hash and _generate_tx_nonce (result.value.to_json() → json.loads(...)) — result.value.blockhash is already a typed string. Remove the roundtrip.

Naming / API clarity

• The AEA base class Helper._generate_tx_nonce is chain-agnostic ("nonce" = replay-defeater). For Solana readers this is confusing — blockhashes are an anti-replay mechanism, distinct from Solana's actual durable-nonce accounts. Add a docstring clarifying what "nonce" means for this plugin, or introduce a Solana-specific name alongside the overriding method.
• generate_tx_nonce(seller, client) (the @staticmethod sha256 variant on SolanaHelper) is literally unrelated to Solana — pure deterministic hash. Document or remove.

Dependency hygiene

• The plugin is pinned tight at solana>=0.33.0,<0.34.0 in #(this PR). solana>=0.36 removed the solana.transaction module (replaced by solders.transaction) — adopting it requires refactoring all Transaction imports + call sites. Worth doing so we can track solana-py's current line.
• anchorpy>=0.20.0,<0.21.0 — anchorpy 0.21+ requires solders>=0.21 + solana>=0.36 (the two bumps above). Group these.
• CI coverage: the plugin has 33 tests but 9 are devnet-network-dependent (hit api.devnet.solana.com faucet) and are failing today due to airdrop rate-limit exhaustion. Either gate them behind a @pytest.mark.integration marker excluded from default runs, or stub the RPC boundary so unit tests remain deterministic.

Mypy / pylint baseline

Not blocking this follow-up, but for reference the plugin currently carries 51 mypy errors and ~20 pylint warnings that long predate any recent change. Worth a tracked clean-up pass.

Context

• Downstream unblock: Re-enable open-aea-ledger-solana (anchorpy/cachetools conflict) open-autonomy#2479
• The minimum-diff dep bump: #(this PR)
• Critique that drove this list: internal review of the inline BlockhashCache substitute, which flagged that faithfully preserving upstream behaviour preserved upstream's pre-existing flaws.