[transfix]

What — H0 (holon prereq) closure

H0 of the holon plan is "close data() object replication (DISTRIBUTED_STATE §52)". A code audit found the core path is already landed and live (the roadmap's "NOT STARTED" was stale): per-shard codec registry → dataChanged→set_data emit (inline / content-addressed blob_ref) → apply_remote inline-decode / blob-hydrate → node.data(), plus snapshot capture+ingest, across all four transports. The existing inproc tests already prove the round-trip, blob-ref, snapshot, and stream-descriptor cases.

This PR closes the small gaps that remained to meet §52's own design + testing bullets:

• S1 — integer data() key fix (state_codec_registry.cpp). register_builtin_codecs keyed integer codecs on the C++ spellings (int32_t…), but state::dataTypeName() returns the app's registered names (int, unsigned int, unsigned short, int64, uint64), so an integer data() silently failed to replicate (encode threw "no codec", swallowed). Added codec aliases under the app spellings. Real bug.
• S2 — value type_name across apply + snapshot (state.{h,cpp}, state_sync_adapter.cpp, state_cluster_shard.cpp). The emit carried valueTypeName() but apply_remote::set_value stamped "std::string" and snapshot() never captured it. Added a signal-free state::valueTypeName(const std::string&) setter and wired apply + snapshot capture/ingest.
• S4 — IPC data() round-trip test (state_transport_ipc_test.cpp): inline object + builtin-int (exercises S1 over a real transport) + blob-ref, closing §52's "both IPC and gRPC" testing bullet (gRPC coverage exists but compiles out by default).
• S5 — set_data echo-suppression test (distributed_state_session_test.cpp): an applied set_data emits no local mutation; a quiescent re-pump moves nothing.
• S6 — roadmap correction (DISTRIBUTED_STATE_ROADMAP.md §52): rewritten from "NOT STARTED" to the landed state + accurately-scoped deferrals, so a future planner doesn't rebuild what exists.

Deferred (documented in §52, not done here)

• format_desc codec on a production shard. The codec exists but no production shard registers it — because none currently replicates /streams descriptors: the conference relay scopes replication to conf.* (enforce_interest), descriptors are rooted at the bare stream token, and cross-node media rides stream_bridge/media_frame, not the state tree. Registering it on the arisrv shard today would be dormant. It belongs with whatever first replicates descriptors (a holon / a conference feature). Covered in isolation by TwoNodeStreamDescriptorDiscovery.
• WebSocket late-joiner blob hydration; the async hydrator late-install guard (must close before any non-blocking chunk transport); the oversized-value() blob resolver.

Testing

All affected suites green: distributed_state_session (20, +3 new), state_transport_ipc (19, +1 new), state_snapshot_protocol (4), state_sync_adapter (8), state_codec_registry (9). New tests: TwoNodeInprocIntegerDataReplication, TwoNodeValueTypeNamePreserved, SetDataApplyDoesNotEcho, StateTransportIpcTest.DataPayloadReplicatesOverIpc (8× repeat clean). Formatted with clang-format 18.1.8.

Context

First step of the holon plan after the clock_driver time-slice framework (#766). Next: H1 — the single holon (shared scene + avatars). Note for H1: replication is TRUST-ALL and origin_node_id is forgeable, so the first physics payloads (pose/body_state) must go through the server-mediated-write / propose→veto→override path, not direct replicated writes — and no codec exists for any physics type yet (authoring one is an H1 dependency).

[transfix]

Ran a 3-lens adversarial review of this branch (S1 codec type-exactness / S2 replication correctness / S4+S5 test validity + hidden regressions), each finding to be independently verified before surfacing. Zero confirmed findings — all three reviewers returned clean (journal verified: genuine {findings: []}, not an empty/errored result).

Specifically checked: the integer codec aliases any_cast the exact registered types (int/unsigned int/unsigned short/boost::int64_t/boost::uint64_t) with no duplicate-key or existing-consumer regression; the signal-free valueTypeName setter doesn't re-enter the journal and the two lock acquisitions in apply_remote aren't nested; the snapshot capture/ingest of type_name is echo-safe and doesn't overwrite on empty; the IPC test genuinely exercises Payload serialization (inline + blob offload, store.size()>0) with no data race on the shared store; and S5's on_local_mutation/pumped2==0 assertions actually prove no re-emit.