Heterogeneous targets · R-1 (research): data-dependent shapes — MoE capacity policy

R-1 — Data-dependent shapes: MoE capacity policy (research)

Data-dependent shapes — MoE routing, top-k, NMS — have no answer, and no bound analysis (P1-1) will produce one, because the extent depends on values, not on input shapes. But production MoE already solves this with a capacity factor + token dropping, and that solution is Vx-shaped: the buffer is known, the occupancy is not, and the overflow behaviour is a policy decision currently buried in a YAML float. Nobody else can express this, because nobody else has capacities in the type system. Genuinely unproven — hence R-1, not P-anything. Do not block release on it; it is the strongest long-term claim in the analysis.

Proposed (research direction)

Memory ExpertBuf {
  within: Memory::HBM, capacity: 8 MB, managed: explicit,
  on_overflow: drop                  // NEW: drop | spill(Memory::HBM) | recompute
}

Diagnostic

E60xx: routing into 'ExpertBuf' has a data-dependent extent and no overflow policy
  help: declare `on_overflow:` on the space, or bound the routing extent

Status

Research; write it down, don't block release. Relates to P1-1 (bounded dynamism handles the ~90% that is shape-derived; this is the data-dependent tail).

From the Heterogeneous Target Gap Analysis (§9). Status tracked there and in the P0 decision log. Sibling to the landed P0 work (P0-1 fa056b4e, P0-3 33ce0b64, P0-4 5e10637a, P1-4a-core 34b31902).

Start with §9 of docs/discussions/heterogeneous_target_gap_analysis.md and compare the tracked status in docs/discussions/heterogeneous_p0_decision_log.md. Treat the proposed ExpertBuf overflow policy and E60xx diagnostic as the research scope; done means the data-dependent-shape policy is documented without blocking release.