feat(fast-inbox): resolve inbox endpoints by message total and preflight checkpoints against the effective parent

[spalladino]

Context

The bucketless Fast Inbox rebuild lets ordinary L2 blocks end at arbitrary message prefixes and resolves a live Inbox bucket only when a checkpoint is completed and published. Two L1 primitives are missing for that: a one-call way to find the newest live bucket boundary at or below a locally known message count, and a publication preflight that checks a checkpoint's header and its final Inbox consumption against the parent propose will actually use at execution time, rather than against a caller-supplied parent total.

This PR adds both as additive contract APIs plus their TypeScript wrappers. Node behaviour is unchanged: the publisher still calls validateHeaderWithAttestations, and the sequencer/validator selection logic is untouched. P3 switches the node over.

Approach

• IInbox.getBucketAtOrBeforeTotal(uint64 upperBound) → (uint64 seq, InboxBucket bucket). Cumulative totals strictly increase per bucket, so the answer is the newest live bucket whose total does not exceed the bound. The live interval is [oldest, current] with oldest = current - ringSize + 1 once the ring wrapped (subtraction first, so current + 1 cannot overflow at the sequence maximum). The search probes the newest four live entries one by one, checking for oldest before every step down, then binary-searches only the still unscanned interval [oldest, lastScanned - 1]. Overwritten ring entries are never dereferenced and genesis is a candidate only while sequence zero is live. Not found (the bound is below the oldest retained total) reverts with Inbox__NoBucketAtOrBeforeTotal(upperBound, oldestLiveTotal).
• IRollup.validateCheckpointHeaderAndInbox(CheckpointPreflightArgs calldata) → uint64 bucketHint. CheckpointPreflightArgs bundles the validateHeaderWithAttestations argument set with expectedTotal and expectedParentCheckpointNumber (the flat ten-argument form hit "stack too deep" and the Rollup has only 725 bytes of runtime headroom left; a single calldata struct keeps the forwarder small). Inside RollupOperationsExtLib: derive the effective parent with STFLib.getEffectivePendingCheckpointNumber(block.timestamp), exactly as propose prunes before validating; require it to equal the claimed parent (Rollup__UnexpectedParentCheckpoint(expected, actual)); run the shared header checks, which compare the header's lastArchiveRoot with that parent's archive; resolve expectedTotal to a live bucket (Rollup__InboxTotalNotAtBucketBoundary(expected, actual) if it is interior) and run ProposeLib.validateInboxConsumption on it with the parent's stored total — the same settlement, monotonicity, cap and censorship predicate propose enforces. The old validateHeaderWithAttestations entry point is untouched.
• Constant alignment. MAX_L1_TO_L2_MSGS_PER_BLOCK is now exported to Solidity through scripts/constants-codegen/solidity.json, and MAX_MSGS_PER_BUCKET aliases the generated constant so the node's completion guard and the Inbox rollover cannot drift apart. A Forge test pins both the alias and the rollover behaviour to the generated value.
• Bindings. Regenerated L1 artifacts; InboxContract.getBucketAtOrBeforeTotal(upperBound) returns { seq, bucket } | undefined (undefined on the not-found revert); RollupContract.validateCheckpointHeaderAndInbox(l1TxUtils, args, { time, stateOverrides, from }) runs the preflight over eth_simulateV1 with a block-time override and the caller's state overrides — the transport SequencerPublisher.validateCheckpointHeader already uses — and returns the bucketHint.

Rollup runtime bytecode is now 24,461 bytes (115 bytes under the limit; it was 725 under before). Gas report and benchmark regenerated.

Tests: InboxBucketSearch.t.sol (genesis/overwritten genesis, exact/interior bounds, full and spilled buckets, a hit on each walkback probe with read-count bounds, fallback immediately beyond the walk excluding the scanned suffix, fewer than four live entries, ring wrap with oldest-entry hit/miss, maximum-sequence arithmetic via harness ring surgery, fuzz against a brute-force reference), CheckpointPreflight.t.sol (preflight vs propose parity: fresh chain, consume nothing, interior total, wrong hash, wrong parent identity, equal/later execution timestamp settlement, censorship, cap escape, published parent, automatic prune, invalidated parent, shared header checks), resolve-by-total cases in ProposeInboxConsumption.t.sol, and anvil-backed unit tests for both TS wrappers.

Stacked on #25413.

Part of A-1928.