Add a UTxO RPC (gRPC) chain producer

[Jimbo4350]

Draft: builds and unit-tests pass, but has never been run against a live cardano-rpc server. See "What is not done" below.

Context

Sieve has had one block source: a local node over node-to-client ChainSync. cardano-rpc now serves the UTxO RPC SyncService, and its FollowTip method — despite the name — replays history from an intersection point, so it can drive a full sync rather than only tail the tip.

This adds it as a second producer, selected by --rpc-socket-path, --rpc-address, or --via-rpc (which derives rpc.sock beside --socket-path, cardano-rpc's own default).

The original design note in Cardano/Sieve.hs always anticipated this: "chain-sync or ogmios or cardano-rpc to get blocks — This needs a wrapper around it to switch between the different sources of blocks."

Shape

New modules under Cardano.Sieve.CardanoRpc, a sibling of Cardano.Sieve.Node rather than a child, because only the transport differs:

module	job
Decode	native_bytes → BlockInMode
Endpoint	where the server is; grapesy connection setup
Response	protobuf projections, so the loop reads as chain logic
Follow	the stream loop

Both producers share sieveBlock, startPoints and all of Node.Insert, so they write identical databases.

The one genuinely new capability

Over ChainSync, cardano-api decodes inside connectToLocalNode and sieve is handed a BlockInMode. Over gRPC the client gets AnyChainBlock.native_bytes and must decode them itself — sieve had no ByteString -> BlockInMode anywhere.

Those bytes are the ChainDB's on-disk serialisation (GetRawBlock), so the consensus disk codec applies, and the era tag it carries makes one decoder work across every era. Taking the raw bytes rather than the parsed protobuf block in the same message is what keeps Node.Decode/Node.Encode reusable, and with them the ledger-exact semantics they encode.

Two things the gRPC path gets for free

A real at-the-tip signal. Every response carries the node's live tip, so "am I at the tip?" is asked and answered rather than inferred from an empty pipeline as collectFlushingWhenIdle must. One comparison drives both the deferred index build and the flush cadence.

A continuous decoder cross-check. Every block arrives with the slot and hash the server derived from the same bytes. The loop compares them on every block, so the decoder proves itself rather than being trusted — which matters while the post-Byron arms have no golden fixtures.

How to trust this PR

cabal build -j4 clean with no warnings; fourmolu applied; all 57 tests pass.

The decoder is tested against real mainnet Byron blocks, borrowed from cardano-rpc's own fixtures. Byron is the era worth testing because EpochSlots is consumed on its decoder arms and nowhere else — the epoch-boundary block landing on slot 2,160,000 is only correct if slots-per-epoch really is 21,600. Trailing bytes, truncation, non-CBOR and an unknown era tag are all asserted to fail.

CLI validation verified by hand:

$ cardano-sieve --rpc-socket-path /tmp/rpc.sock --database x --until 100
cardano-sieve: --until is not supported when syncing over UTxO RPC

$ cardano-sieve --via-rpc --database x
cardano-sieve: --via-rpc needs --socket-path to derive the rpc socket path

What is not done

• Never run against a live server. No block has actually travelled this path.
• The differential test is the real acceptance test and has not been run: sync the same range over both producers and diff outputs/unspent/blocks. Anything but identical counts means the port is wrong.
• Post-Byron decoder arms have no fixtures — covered only by the runtime cross-check. Capturing one Conway block's native_bytes would close this.
• Rollback is untested. Bounded historical ranges cannot exercise it, same gap the ChainSync path has.
• --until is rejected over gRPC (FollowTip has no stopping condition), so the bench scripts cannot yet compare the two producers.

Incidental

test/Main.hs did not compile against the pinned cardano-api — AdaAssetId is now a constructor of AssetId rather than a standalone export. Fixed here because nothing could be tested otherwise; happy to split it out.

sieveBlock, startPoints and describeSelectors become exported from Node.Fetch rather than duplicated — each encodes a policy that must not differ between producers. No logic in Fetch.hs moved.

🤖 Generated with Claude Code