feat(fast-inbox): flip consensus to streaming inbox (A-1384)

[spalladino]

Coordinated cutover of the Fast Inbox (AZIP-22) project (issue A-1384, FI-14). Flips the authoritative path across L1, circuits, and the node from the legacy inHash/frontier-tree consumption to the streaming rolling-hash Inbox.

Stacked on the node phase: #24784 (A-1379) -> #24785 (A-1380) -> #24786 (A-1381) -> #24787 (A-1382) -> #24788 (A-1383). Base is spl/a-1383-validator-streaming.

Merge gate

• A-1431 (domain-separation decision) must resolve before this merges. It is deliberately deferred and nothing is adopted here. If it adopts a separator, the rolling-hash link-encoding change must be inserted below this PR in the stack.

L1 (l1-contracts)

• ProposeLib.propose calls validateInboxConsumption against the parent checkpoint's consumed position (read from the parent temp-log record) instead of inbox.consume() + the Rollup__InvalidInHash check; the returned consumed total is stored in the new record.
• ProposeArgs gains an unsigned bucketHint calldata field (out of the attested payload digest).
• TempCheckpointLog / CompressedTempCheckpointLog carry {inboxRollingHash, inboxMsgTotal, inboxConsumedBucket}; genesis is {0,0,0}.
• EpochProofLib.getEpochProofPublicInputs anchors the rolling-hash chain start to the record of checkpoint start - 1 (Rollup__InvalidPreviousInboxRollingHash).
• Inbox.sendL2Message returns and emits the compact cumulative message index.

Circuits (noir-projects)

• MAX_L1_TO_L2_MSGS_PER_BLOCK 1024 -> 256.
• L1ToL2MessageBundle drops num_real_msgs; a single num_msgs drives the compact tree append and the message-sponge absorb.
• SpongeBlob::absorb_block_end_data absorbs the L1-to-L2 tree root for every block.

Node (yarn-project)

• Streaming is the only path: the streamingInbox flag is removed across foundation/stdlib config + the sequencer/validator plumbing.
• Sequencer: sources the parent bucket from the fork's L1-to-L2 leaf count + getInboxBucketByTotalMsgCount (cross-checkpoint); feeds inHash zero; relaxes waitForMinTxs for message-only blocks; threads the consumed bucket seq as the propose bucketHint. Automine mirrors the selection.
• Validator: per-block acceptance + checkpoint re-execution always run; the checkpoint rebuild derives the consumed message list from the buckets between the parent checkpoint's position and the last block.
• World state: appendL1ToL2MessagesToTree / handleL2BlockAndMessages append real leaves at compact indices; the synchronizer derives each block's bundle from its leaf-index range.
• Blob / constants / bundle: per-block blob root, MAX_L1_TO_L2_MSGS_PER_BLOCK = 256, L1ToL2MessageBundle drops numRealMsgs; provable per-checkpoint ceiling 2457 -> 2234.
• Archiver / TXE: the dead inHash cross-check is removed; TXE appends unpadded compact leaves.
• L1 ABI: @aztec/l1-artifacts regenerated for the new ProposeArgs (gitignored generated output; the ethereum package builds against it).

Verification

• L1 forge test: full propose/inbox suite green (ProposeInboxConsumption, Inbox, InboxBuckets, Rollup incl. a new anchoring negative, RollupFieldRange, FeeRollup, ValidatorSelection, escape-hatch, tmnt207/419).
• Circuits nargo test: parity, block_root structure, msgs_only green on touched crates.
• Node jest (runnable locally): validator proposal_handler (34) + streaming checks (14), sequencer inbox selector (9), world-state native (56), blob-lib + stdlib. tsc for the broken-build-zone packages (prover-client, sequencer-client, validator-client, etc.) is CI's job of record.
• Codex (gpt-5.6-terra) reviewed both the L1/circuits layer and the node wiring: no consensus-critical issue; bucket cursor/range (fromExclusive, toInclusive] and pipelining-parent sourcing verified correct.

VK regen and Prover.toml sample-input regen ride CI (local bb write_vk OOMs on rollup_root).

Node prover + archiver keying (completed in this PR)

• Prover per-block message split: the checkpoint's messages are sliced per block from the blocks' L1-to-L2 leaf-count ranges; each block root appends its own real slice at compact indices with per-block start/end snapshots and a full-height frontier hint. startCheckpoint no longer inserts messages up front.
• Checkpoint keying from records: getL1ToL2MessageCheckpoint binary-searches the block records for the first block whose L1-to-L2 leaf count exceeds the message index (portal claim helpers depend on this); the prover-node derives the checkpoint's consumed messages from the Inbox buckets.

Codex (gpt-5.6-terra) reviewed the prover/keying rework: no consensus defect; slice partitioning, bucket-range agreement, fork ordering, frontier height, and the checkpoint boundary rule all verified.

Left for FI-18 cleanup

The message_store InboxLeaf index-formula methods (smallestIndexForCheckpoint etc.) have no correctness-critical live callers post-flip; the public-simulator's next-checkpoint fetch degrades safely without them (simulates without the not-yet-consumed messages). Deleted in FI-18.

Accepted residual risk

A-1390 (unconsumed-bucket overwrite protection) stays post-flip: a consumption backlog older than the Inbox ring can overwrite unconsumed buckets and halt proposals until an upgrade. Acceptable for non-production lines.

Review follow-up (phase-2 final review)

A final review pass appended fix(fast-inbox): thread per-block message sponges and wire the msgs-only block root in the prover:

• The prover supplied the checkpoint-wide message sponge as every non-first block root's start sponge, which the block-merge/checkpoint-root continuity asserts reject whenever a non-first block carries messages. The sponge is now threaded per block.
• A zero-tx non-first block (the message-only shape this PR lets the sequencer produce) was rejected by BlockProvingState and the lightweight builder, and the orchestrator never selected the msgs-only block root. Both now accept it and route it through BlockRootMsgsOnlyRollupPrivateInputs.
• The required bucketHint parameter added to enqueueProposeCheckpoint here is now passed by the publisher unit/integration tests and the e2e synching test (the integration suite's consumption model is reworked for streaming semantics in chore(fast-inbox): delete legacy node message paths (A-1388) #24793).