Measured 2026-08-31 on the released 5090 container (binary 7eb66643, one shared-prefix slot configured). Five /v1/messages requests sharing an identical ~83.5K-token system block with cache_control: {"type": "ephemeral"} on the system text block all classified reuse=root, prefix_cache_hit_tokens=0 - seed and every fan cost ~34 s each. The schema parser does record system cache boundaries (anthropic_schema.cpp flatten_system_value), so either the boundary never becomes a SharedStablePrefix capture group, shared capture never runs at completion, or the follow-up shortlist never matches.
Receipt context: omp-ninfer docs/measurements/2026-08-31-fanout-probe-main.json (the Responses-path fanout in the same session: branch-from-older-id is always root as well). Together these mean no same-lane fanout path reuses a shared base on the released build; the v0.4.x engine train (upstream 099d9032 value-aware shared scheduling, 59febed prefix retention, plus Responses-path markers) is the fix vehicle. Diagnosis starting point: trace a marked anthropic request through frontend capture-group construction to the shared catalog and the next request's shortlist.
Measured 2026-08-31 on the released 5090 container (binary 7eb66643, one shared-prefix slot configured). Five /v1/messages requests sharing an identical ~83.5K-token system block with
cache_control: {"type": "ephemeral"}on the system text block all classifiedreuse=root, prefix_cache_hit_tokens=0- seed and every fan cost ~34 s each. The schema parser does record system cache boundaries (anthropic_schema.cpp flatten_system_value), so either the boundary never becomes a SharedStablePrefix capture group, shared capture never runs at completion, or the follow-up shortlist never matches.Receipt context: omp-ninfer docs/measurements/2026-08-31-fanout-probe-main.json (the Responses-path fanout in the same session: branch-from-older-id is always root as well). Together these mean no same-lane fanout path reuses a shared base on the released build; the v0.4.x engine train (upstream 099d9032 value-aware shared scheduling, 59febed prefix retention, plus Responses-path markers) is the fix vehicle. Diagnosis starting point: trace a marked anthropic request through frontend capture-group construction to the shared catalog and the next request's shortlist.