feat: world-state per-block message insertion (A-1380)

[spalladino]

Part of the Fast Inbox stack (AZIP-22, FI-10 / A-1380). Base: `spl/a-1379-archiver-buckets`.

What this changes

• `NativeWorldStateService.handleL2BlockAndMessages` no longer throws when a non-first block carries an L1-to-L2 message bundle. Any block may now transition the L1-to-L2 message tree.
• First-in-checkpoint bundles are still padded to `NUMBER_OF_L1_L2_MESSAGES_PER_ROLLUP` (so the legacy call shape produces bit-identical trees); non-first bundles are appended exactly as given (the post-flip compact append the circuits perform). Padding stays in this method as a transitional concern and moves to the caller at the flip.
• The legacy "non-first blocks carry no messages" rule moves from a throw inside world-state to a transitional assertion owned by the synchronizer call site (`handleL2Block`). The synchronizer's per-checkpoint message fetch is unchanged pre-flip.
• `MerkleTreeAdminDatabase.handleL2BlockAndMessages` JSDoc updated to document the per-block semantics.

C++ investigation outcome (no barretenberg change needed)

Investigated the native sync/unwind paths under `barretenberg/cpp/src/barretenberg/world_state/`. Per-block message insertion and per-block unwind already work with no C++ change:

• The L1-to-L2 message tree is the same append-only `FrTree`/`FrStore` as the note-hash tree, which already varies its leaf count per block and unwinds per block.
• `commit_block` increments the tree's block height and writes a per-block `BlockPayload` (`{size, root}`) on every `sync_block`, even when zero leaves are appended, so the message tree's per-block history stays in lockstep with the other trees.
• `unwind_block` reverts exactly one block's appended leaves from that stored per-block metadata, generic over leaf count. C++ has no concept of an L2 checkpoint or "first block" — that is a TS-only invariant — so a mid-checkpoint block that appended message leaves unwinds correctly.

The TS-side pending-chain rollback (`handleChainPruned` -> `unwindBlocks`) just unwinds to a block number and carries no checkpoint assumption about the message tree either.

Tests

• Every-block equivalence: a legacy-shaped multi-block checkpoint leaves the message tree byte-identical (root + size) on every non-first block, while the archive advances per block — proving the degenerate path is a no-op at every block, not just at checkpoint ends.
• Per-block appends: a checkpoint whose blocks advance the message tree by exactly their (unpadded) bundle size on non-first blocks; forks opened mid-chain see exactly the bundles up to that block.
• Per-block unwind: unwinding a message-carrying non-first block reverts exactly its appended leaves back to the pre-block state, and re-syncing afterwards reconverges.
• Synchronizer guard: a non-first block carrying messages is rejected by the call-site assertion.

world-state type-checks clean and all world-state unit tests pass (76 across the two modified suites). The full `yarn build` still fails only in the pre-existing stale-artifact packages (noir-protocol-circuits-types, simulator, prover-client), unrelated to this change.

Review follow-up

Codex (routine review) confirmed the C++ no-change conclusion and the pre-flip synchronizer soundness, and flagged that the TXE world-state caller (txe/src/state_machine/synchronizer.ts) padded the bundle unconditionally. TXE mines one block per checkpoint (always the first block), so this was a no-op in practice, but the second commit makes its padding conditional on indexWithinCheckpoint === 0 to keep the caller aligned with the new per-block semantics.