feat(core): add rfq_swap_replacer built-in post-processor

[carloszanella]

Summary: PR body describes the feature and the important decisions well. I made Codex write it because Claude writes like a regard. Tested locally, the post processing takes basically no time. To enable liquorice we need to add it to the post processor protocol list and disable it from the solver so the quotes are not used partially.

Business logic

Feature

The new built-in post-processor rfq_swap_replacer runs when a worker pool lists it in its post-processors. It replaces a swap, a segment between two route tokens, or the whole order with RFQ swaps from components that the market tracks, using both direct RFQ components (rfq:) and RFQ components through the fallback router (fallback:rfq:). It keeps a change only when the net amount out after gas beats that of the route it received. Worker pools with no post-processors do not change; worker pools with other post-processors get only the failure fallback described below.

• The new worker pool key algorithm_exclude_protocols removes protocols from the algorithm's graph and its fallback pricing, while keeping them visible to post-processors. For example, ["rfq:bebop"] means Bebop enters a route only through the replacer, so the replacer's gain is measured alone. The existing key exclude_protocols keeps its meaning: the worker pool never routes through those protocols.

Business decisions

• Partial fills. An RFQ swap can replace a swap in the middle of a route only if the protocol accepts a smaller amount in than it quoted. The code keeps an allow-list in PARTIAL_FILL_PROTOCOLS (fynd-core/src/post_process/rfq.rs), which currently holds rfq:bebop and rfq:hashflow. Every other RFQ protocol can replace only the whole route or swaps that start at the sell token, because only those swaps receive the exact amount in that the quote was for.
• Failures. This fallback applies to every worker pool with post-processors. If a processed route fails the worker's checks, including the fallback pricing of a fallback: leg, the worker returns the algorithm's route instead of a failed quote. It logs a warning with the worker pool, the algorithm, the post-processors that made the route, and the error. It also increments worker_pool_post_processed_route_rejected_total{pool}.
• Deadlines. The search stops at the post-process deadline, 10 ms before the task deadline. The replacer checks the deadline before each candidate, but a simulation that has started always runs to the end.
• Prices. The replacer uses indicative RFQ state. The binding quote at encode time can be worse or can fail, so a replaced route can fill worse than the algorithm's route or fail at encoding.
• Unchanged routes and skipped components. The replacer leaves routes unchanged when they have zero gas or the market read fails. It skips components that the request excludes or the worker pool's drop rule removes.
• Reserved name. An embedder cannot register a custom post-processor with the reserved name rfq_swap_replacer.
• Public getter. The new public getter is OrderQuote::post_processed_by(). The field uses #[serde(skip)], so the HTTP response and the OpenAPI do not change.

Observability

• worker_pool_post_processor_outcomes_total{pool,post_processor,outcome} records the outcomes replaced, kept, timed_out, and skipped.
• worker_pool_post_processor_gain_bps{pool,post_processor} is a histogram of the gain with its own buckets.
• worker_router_best_quote_pool, worker_router_won_volume_gas_token and worker_router_unpriced_won_quotes_total gain the label post_processed_by. Its value is the post-processor names joined with +, or none. Existing series split by this label; a sum by pool gives the old totals.
• worker_pool_post_processed_route_rejected_total{pool} increments when a processed route fails the worker's checks.
• The warning log records the worker pool, the algorithm, the post-processors that made the rejected route, and the error.

🤖 Generated with Claude Code

[Troshchk]

Thank you!
Left some comments and 2 qs:

1. We only support rfq: now, should we insert fallback:rfq instead?
2. Can we somehow limit which rfqs are used in per hop swap(with partial fills supported) and whole route swap(partial fills do not matter

[Troshchk]

Thank you! A couple of qs

[Troshchk]

Thank you!