comm: the lane lease as a communicator split

This maps three shipped fak coordination surfaces onto the MPI communicator primitives they are shaped like, so the analogy is documented in one place instead of inferred from terse inline comments. It is part of the MPI-shaped message-passing epic (#639).

Honesty caveat (read first). This is structural analogy, not a message-passing transport. A lane lease is a coordination decision enforced by refusal (two overlapping leases serialize; one is denied), deterministic and dos-verifiable — no collective moves bytes between agents at this layer. topobench’s “savings” are MEASURED replay token counts capped at the corpus divergence frontier, never an HPC throughput/latency number. And the agent-layer rank/size here is not the tensor-layer rank/size of internal/model’s DistComm serving collective — see “What this is NOT”.

---

1. dos arbitrate + dos.toml [lanes] ≈ MPI_Comm_split

MPI_Comm_split(comm, color, key) partitions a communicator: every rank passing the same color lands in the same sub-communicator, disjoint from the others. fak’s lane lease is the same partition, decided by admission rather than rendezvous:

• color = the lane. dos.toml [lanes] (dos.toml:32) declares the lane taxonomy, and [lanes.trees] (dos.toml:135) gives each lane its canonical, prefix-disjoint file tree. A worker requests a lane; dos arbitrate admits it iff its tree is disjoint from every live lease (PARTITION.md: “one lease per tree”).
• the split is enforced by refusal. Where MPI_Comm_split places a rank, the arbiter admits or denies one: two workers whose trees intersect cannot both hold an exclusive lease, so they serialize (one waits) instead of colliding. The partition is a decision the kernel can refuse, not a barrier all ranks must reach — which is why a single absent worker never deadlocks the others.
• deterministic and verifiable. The admission is a pure function of the requested lane, mode, tree, and the live-lease set (the dos arbitrate kernel), so the same inputs yield the same verdict — dos-checkable, not a timing-dependent race.

The load-bearing difference from MPI_Comm_split: a communicator split is a collective all ranks call together; a lane lease is an asymmetric admission one worker requests and the arbiter grants or refuses. No bytes cross between the workers in the same lane — the lease coordinates who may write which files, it does not transport a message.

2. abi.ShareScope ≈ the communicator isolation scope

abi.ShareScope (internal/abi/types.go:93) is the CLOSED, additive isolation scope a shared resource (a Ref) carries — the analogue of which communicator a name is visible within:

ShareScope	meaning	communicator analogue
ScopeAgent	private to one agent (the fail-closed default)	a rank-private buffer (no communicator)
ScopeFleet	shareable across the fleet’s trusted partition	the fleet communicator
ScopeTenant	shareable within a tenant boundary	a tenant sub-communicator

The default ScopeAgent is fail-closed (private): a value is shared only when its scope is explicitly widened, the same discipline that keeps a name from leaking across a communicator boundary it was never published to. This is visibility scope, not a broadcast — widening a scope authorizes a later share, it does not move data.

3. topobench TopologyGenome ≈ MPI_Cart_create / MPI_Graph_create — but it SEARCHES

MPI_Cart_create / MPI_Graph_create impose a topology a caller declares. fak has both halves, and the distinction is the point of this section:

• DECLARE — internal/agenttopo. agenttopo declares a named, validated DAG over a comm.Group: who may hand a result to whom, every edge endpoint checked against the group, cycles refused, declaration order preserved (internal/agenttopo/doc.go). This is the direct MPI_Graph_create analogue: a fixed adjacency the caller asserts.
• SEARCH — cmd/topobench + turnbench.TopologyGenome. topobench does NOT declare a shape; it optimizes an anonymous one. A TopologyGenome is three searchable levers — Width (fan-out), SubTurns (orchestrator→worker depth), Lanes (the leaf-lane partition the workers serialize on) (internal/turnbench/toposearch.go:92) — and the search ranks genomes by CreditedSavingsTokens, the measured prefix-reuse + dedup token saving capped at the corpus frontier width (min(Width, Wmax)), so an extrapolated topology can never be crowned by a saving the corpus never recorded.

So: agenttopo is “here is my graph” (MPI_Graph_create); topobench is “find me a good graph, and prove the win is measured, not extrapolated” — the search counterpart to a declared topology. The “savings” are token counts off RunFanoutCell’s measured halves, capped at the divergence frontier — not an HPC metric.

---

What this is NOT

• No bytes move at this layer. A lane lease serializes writers; a ShareScope authorizes a later share; a TopologyGenome ranks fleet shapes by measured token savings. None of the three transports a message between agents — there is no collective moving data here, unlike a real MPI_Bcast/MPI_Allreduce.
• Agent rank/size ≠ DistComm rank/size. internal/model’s DistComm (the serving collective in internal/model/dist_collective.go) carries real cross-process host float32 between devices; its rank/size index GPUs/workers in a tensor collective. The agent-layer “rank” here indexes agents/roles in a coordination graph. They are different ranks over different objects — do not conflate the two.
• pipelinegen is the tensor/device layer, not this one. cmd/pipelinegen runs fak’s NATIVE engine across pipeline-parallel transformer stages (internal/model/pipeline.go): hidden state crossing a serialize→bytes→deserialize boundary between layer-bands, the stand-in for a real NCCL/RPC worker hop. That is a model-internal serving demonstrator (device collectives), not an agent-topology or communicator-split surface. The agent-layer mappings above (agenttopo/topobench/lane lease) are unrelated to it; they inherit no HPC latency, throughput, or progress guarantees from the tensor layer.

---

See also

• dos.toml — the [lanes] taxonomy and [lanes.trees] partition the arbiter splits on.
• PARTITION.md — “one lease per tree”: the disjoint-tree rule behind the split.
• internal/agenttopo — the DECLARED agent-topology DAG (the MPI_Graph_create half).
• internal/turnbench/toposearch.go — TopologyGenome / TopologyFitness, the SEARCH half.
• collectives.md — the canonical MPI-collective map (Combine reduce / DistComm allreduce / ShareScope bcast) and the agent-vs-tensor honesty line this doc’s rank/size caveat points at.
• vdso-revoke-as-comm-revoke.md and proofs/async-addressing.md — sibling MPI-analogue docs in the same epic.