feat(router): true exact-out (in-given-out) routing + regression tests

[JohnnyWyles]

Summary

Replaces the inversion-based GetOptimalQuoteInGivenOut with a direct in-given-out routing path, and adds the regression coverage for it.

SQS previously implemented exact-out quotes by inverting an out-given-in quote. That caused three confirmed production issues:

1. Understated input cost. amount_in excluded the taker fee and spread on the input side.
2. Wrong route selection. Candidates were found in the wrong direction (confirmed ~$1.30/SOL worse on SOL to USDC, routing 80% through a 0.5% spread pool vs 0.05% for exact-in).
3. Corrupted fee metadata. Pool taker fees differed between directions for the same pool (pool 1925 reported taker_fee=0.000 in exact-in and taker_fee=0.001 in exact-out).

What changed

Wiring (add90084, already on the branch):

• GetOptimalQuoteInGivenOut now calls computeAndRankRoutesByDirectQuoteInGivenOut directly instead of inverting an out-given-in quote.
• quoteExactAmountOut.PrepareResult computes amount_in, effective fee, price impact, and spot price from the true exact-out route data (PrepareResultPoolsInGivenOut) rather than the embedded inverted quote. Price impact is (effectiveOutPerIn / spotOutPerIn) - 1, negative when adverse. The legacy inversion branch is retained as a fallback while quoteExactAmountIn is non-nil.

Split-route optimization for exact-out:

• getSplitQuoteInGivenOut is the input-minimising dual of the exact-in getSplitQuote. Where exact-in splits the input across routes to maximise output, this splits the desired output across routes to minimise the total input required, via a dynamic program with inverted objective (unreachable states seeded with a nil sentinel rather than zero; selection keeps the smaller input; per-increment cost uses CalculateTokenInByTokenOut and charges exact-out taker fees).
• The truncation remainder from per-route output proportions is assigned to the last selected route, and each route's input is computed from its exact (remainder-adjusted) output, so the split outputs sum to exactly the requested tokenOut.
• GetOptimalQuoteInGivenOut compares the single best route against the split quote and returns whichever requires less input, mirroring the out-given-in single-vs-split selection.

Tests (this commit):
The wired path only had coverage for the legacy inversion fallback (where the embedded quoteExactAmountIn is set). The true path (nil quoteExactAmountIn) was untested. Added:

• NewExactAmountOutTrueQuote test helper that builds a quote on the true path.
• TestPrepareResult_ExactOut_TruePath_EffectiveFee: per-route compounded taker fee weighted by output share.
• TestPrepareResult_ExactOut_TruePath_PriceImpact: re-derives price impact from the quote's own reported spot price and asserts the adverse-trade negative sign convention.
• TestPrepareResult_ExactOut_TruePath_TakerFeeMetadata: zero-taker-fee pool reports zero effective fee; a non-zero pool reports exactly its fee (pool 1925 metadata regression).
• TestGetOptimalQuoteInGivenOut_Mainnet_FeeCorrectness: end to end against mainnet state, asserts positive fee-inclusive amount_in, in-given-out route orientation, and an effective fee matching the compounding of the chosen pools' taker fees.

Verification

CI is the verification gate for the test run: the suite cannot compile on native Windows (wasmd needs the cgo NewKeeper, and wasmvm ships no Windows library), so these were authored and gofmt-checked locally but executed only in CI (Linux, CGO enabled).

Pass criteria for the wiring, to confirm on a preview/live node:

• amount_in for exact-out is close to the equivalent exact-in input (was understated by the inversion).
• price_impact for exact-out is negative and similar magnitude to exact-in.
• Route split for exact-out picks the low-spread pool, consistent with exact-in.
• Pool taker fees are consistent between exact-in and exact-out responses.

Scope and behaviour notes

• Exact-out picks the lower-input of single route vs split. The in-given-out path now runs split-route optimisation and returns whichever of the single best route or the best split requires less input for the desired output. A pair resolves to one route when splitting does not reduce the required input, and to several when it does. The new split optimiser is guarded by a test asserting the split never requires more input than the best single route.
• Updated TestGetOptimalQuoteExactAmounOut_Mainnet expected route counts. The previous expectations reflected the old inversion implementation (which borrowed exact-in's split routing in the wrong trade direction). They are set to the counts produced by true exact-out routing.
• Empty input denom on exact-out result quotes. On the true exact-out path, GetAmountIn().Denom and route.GetTokenInDenom() come back empty (the result struct carries amounts and pools but not the input denom string). The new mainnet test asserts on amounts and per-pool taker fees rather than the input denom. Worth a look during review to decide whether the input denom should be threaded into the result.
• CI observability. This branch also tweaks .github/workflows/build.yml to tee go test output to the console (with set -o pipefail) and upload the coverage artifact if: always(), so a failing test surfaces its assertion in the run log instead of being swallowed into an unuploaded report.json.