test(e2e): pin gas_estimation public-payment txs to one block to deflake fee comparison

[spalladino]

Problem

The single-node/fees/gas_estimation test estimates gas with public payment method flakes (CI hash 8b1cc7ba303cf26f, FLAKED-then-retried-green on this branch). The first-attempt failure is the fee comparison at line 200:

expect(received).toBeLessThan(expected)
Expected: < 1044344170000000n
Received:   4341054744900000n

Root cause (grounded in logs + source)

The test sends two token-transfer txs concurrently via sendTransfers: withEstimate uses tight estimated teardown gas limits, withoutEstimate uses the large default limits. It expects withEstimate to pay less, because unused teardown gas is not refunded. transactionFee scales with the landing block's feePerL2Gas.

Under the pipelining preset (minTxsPerBlock: 0, two blocks per slot) the two txs are not pinned to a block, so they land in different blocks:

• withEstimate (0x1a3c0ed…, billed l2Gas=979281) -> block 18, feePerL2Gas=4432900000 -> txFee=4341054744900000 (the failing Received).
• withoutEstimate (0x0461a777…, billed l2Gas=1702550) -> block 19, feePerL2Gas=613400000 -> txFee=1044344170000000 (the Expected < bound).

On a freshly-deployed chain the L1 base fee steps down across checkpoint boundaries (here l1BaseFee 10721937 -> 1462645, with congestionCost=0 and ethPerFeeAsset constant), so the two blocks' feePerL2Gas differ ~7.2x. That price gap swamps the teardown-refund difference and flips the comparison. A passing run had the same tx split but a flat price across the boundary, so it held by luck. The L2 gas price derives from l1BaseFee in sequencer-client/src/global_variable_builder/fee_predictor.ts (computeManaMinFee).

Fix

Mirror the sibling estimates gas with Fee Juice payment method test in the same file: set minTxsPerBlock: 2, maxTxsPerBlock: 2 and wait for the sequencer to go idle before sending, so both txs land in the same block and share one feePerL2Gas. This isolates the teardown-refund difference the test actually verifies.

• No behavioral assertion is relaxed or removed.
• The sibling's totalManaUsed === gasLimits.l2Gas * 2 assertion is deliberately not copied: the two public-payment txs bill different mana by design.
• Added an explicit expect(withEstimate.blockNumber).toEqual(withoutEstimate.blockNumber) so a future batching regression fails clearly rather than resurfacing as a fee-comparison flake.

Verification

• yarn build passes (whole project).
• Red/green could not be reproduced locally: the failure requires the L1-base-fee step to fall between the two tx blocks, which is nondeterministic startup timing (the passing run shows it depends on chance). The green direction is structurally guaranteed by the same-block pin and is the same idiom the non-flaky Fee Juice sibling already uses; the assertion arithmetic was verified to hold at every observed feePerL2Gas.