WIP arc

[SozinM]

📝 Summary

💡 Motivation and Context

---

✅ I have completed the following steps:

• Run make lint
• Run make test
• Added tests (if applicable)

[claude]

Claude finished @SozinM's task in 3m 40s —— View job

---

PR Review: WIP arc

Reviewed the diff against the base branch. The chain-abstraction approach (crates/rbuilder/src/chain.rs compile-time swap, #[cfg(feature = "arc")] gating) is clean and keeps the Ethereum path untouched. Comments below are the issues worth attention; I posted 3 inline.

Blocking / high priority

• Workspace no longer builds without an external arc-node checkout. The arc deps in Cargo.toml (L194–201) are relative path deps into ../../arc-node/, and crates/arc-rbuilder is added to default-members (L31). A fresh clone running cargo check / make lint / make test will fail for any contributor or CI runner without that sibling checkout. The arc feature itself is correctly optional — but default-members undoes that isolation. Recommend dropping arc-rbuilder from default-members (keep it in members) and/or pinning git revs instead of relative paths. (inline)

Correctness to verify

• Empty-requests hash on Prague-active Arc blocks (building/mod.rs L955–959): the arc branch produces Some(Requests::default()), which emits a requests_hash in the header. If arc-node's assembler leaves it None, headers won't be byte-identical and blocks get rejected. Worth a regression test pinning this field. (inline)
• expected_block_gas_limit / post_block_call swallow ProtocolConfig read errors (arc_support.rs) and fall back to defaults. This mirrors arc-node, but a divergence here silently produces a block with the wrong gas limit / base fee that the node will reject rather than fail loudly. Acceptable if intentional; flagging for awareness.

Performance

• Sink mutex held across finalize_block (engine_payload_sink.rs L136–163): the lock spans state-root computation, so multiple builder threads serialize on finalization. For the sub-second TimingsConfig::arc() path this caps finalization to one block at a time. Consider finalizing outside the lock and re-checking best_value on commit. (inline)

Minor

• block_output/mod.rs: pub mod engine_payload_sink; breaks the alphabetical module ordering (inserted before bidding_service_interface).
• EnginePayloadSink ignores built_block_cache (cross-builder coordination cache) — fine if Arc builders don't use it, but worth a one-line note.
• PR checklist (make lint, make test, tests) is unchecked — given the default-members issue above, please confirm these pass on a clean checkout.

Note: I confirmed gas_used passed to post_block_call includes the payout tx (payout is inserted in finalize_block_execution before partial_block.finalize → process_requests), and that merge_transitions runs after the post-block state write — both correct. Passing the full true_block_value as the payout value is consistent with the existing no_competition_bid pattern.

[claude]

Workspace becomes unbuildable without an external checkout. These are path deps into a sibling ../../arc-node checkout, and crates/arc-rbuilder (which depends on them unconditionally) is in default-members (line 31). That means a plain cargo build / cargo check / make lint / make test at the repo root will fail for anyone — and CI — that doesn't have arc-node checked out at exactly that relative path. The arc feature on the rbuilder crate is properly optional, but adding arc-rbuilder to default-members defeats that isolation.

Consider one of:

• Remove arc-rbuilder from default-members (keep it in members), so default builds skip it.
• Gate the whole arc workspace behind an env/profile, or use git deps with a pinned rev instead of relative path deps.

Either way, please confirm cargo check on a fresh clone (no arc-node sibling) still succeeds before merge.

[claude]

The state mutex is held across the entire finalize_block call, which includes state-root computation — the most expensive step. With multiple building-algorithm threads feeding this sink and sub-second Arc blocks (TimingsConfig::arc), every thread serializes on this lock during finalization, so effective finalization throughput is single-threaded.

The cheap best_value comparison must stay under the lock, but consider releasing it before finalize_block and re-acquiring to commit best_value (re-checking the bound after finalize, since another thread may have advanced it). local_ctx would need to move out of the shared state for that. Not blocking, but worth measuring for the fast-block path.

[claude]

When Prague is active this yields Some(Requests::default()) (empty), producing a non-None requests_hash in the header (hash of an empty requests list). If arc-node's ArcBlockAssembler omits requests_hash entirely (sets it to None) for these blocks, the header will not be byte-identical and the block will be rejected. Please confirm arc-node sets the empty-requests hash here too (you mention E2E byte-identical verification — worth a regression test pinning this specific field across a Prague-active Arc block).