5.1 Protocol work not implemented in RTL¶

Lossless partial representation¶

Protocol 1.0 assigns FP32 to switch-to-switch partial DATA. This rounds a local Kulisch result at every switch boundary. A future exact tree profile needs a frozen representation, byte order, chunk geometry, exceptional-value policy, capability record, parser, packetizer, CCL codec, PCAP decoder, and Wireshark fields before receiving a numeric registry value.

Hop reliability¶

PROGRESS, selective acknowledgements, replay identities, retry timing, and window semantics are documented and modeled in software but have no current RTL engine. An implementation must bound bitmap/replay storage, define acceptance and release events, handle duplicates idempotently, and prove forward progress under backpressure and packet loss.

Abort and release draining¶

Ingress admission stops accepting DATA as soon as an Operation is no longer ACTIVE, but the current RTL does not broadcast OP_ABORT or OP_RELEASE into tile FIFOs and range contexts that already accepted work. A production implementation must cancel or drain that tagged work, reclaim incomplete ranges, and enter FREE only after the internal retirement point. Until that mechanism exists, a controller must not reuse an Operation slot while accepted DATA from its previous occupant can still be resident in the datapath.

Arithmetic status¶

The laboratory ACCUMULATION_INEXACT bit reports insufficient accumulator capacity. Further exception bits must not be assigned without defining sticky scope, partial/result propagation, rechunking, old-receiver behavior, and endpoint observability.

Additional collectives and operators¶

An enum value alone is not a specification. Every new operator requires exact input/output semantics, ordering, rounding, exceptional values, state bound, capability compatibility, and bit-exact reference tests. A dot product, for example, requires the product-width Kulisch bound rather than the scalar SUM bound used by the current implementation.

Authentication and distributed recovery¶

Authentication profiles, key ownership, topology-wide rollback, observability, and fault containment remain outside the current implementation. They should be designed together with their trust boundary and resource-exhaustion policy, not added as isolated flags.