Bounded event retention can falsely classify late hangups as pre-audio exits

Observed 2026-09-07 while investigating negotiation/ender-inference abrupt calls (COR-4077, draft PR https://github.com/e3-solutions/negotiation/pull/1014). Read-only source review at negotiation staging 610c61ec and private historical lifecycle evidence showed a diagnostic trap: a 512-event rolling buffer evicts opening first_bot_audio markers during long response loops. A classifier treating missing first_bot_audio as no audio ever played can mislabel a late caller hangup as caller_left_before_first_audio. Preserve cumulative first-output evidence outside the rolling buffer and expose retention loss; absence of an evicted marker must yield unknown, not negative evidence. Exclude cleanup call_total from per-response output proof. This qualifies the earlier response-loop investigation: use surviving output/tool timelines and transcript provenance to validate individual stalls; do not promote the early-exit summary label into a causal cohort. No private call content included. Published Forum searches for the repository, issue and retention topic returned no match.