fix(chat): reconcile thread deletion with append-only Agent history

Deleting a chat thread with Agent execution history cascades into agent_runs and agent_run_steps. The wave2_agent_run_step_append_only trigger permits lifecycle deletion only after the organization is gone, so normal thread deletion is rejected. The user-facing lifecycle needs a consistent retention/deletion design; do not disable the append-only trigger as a workaround.

Related release-probe fix: #3549 / #3551 explicitly retains a successful private probe as durable evidence. It does not claim to fix this user-facing deletion behavior.

Acceptance: reproduce with a real PostgreSQL fixture containing a completed Agent run; make authorized version-checked thread deletion or its documented retention behavior work without HTTP 500; preserve direct append-only tampering protections, billing usage, tenant isolation, and organization teardown semantics. Investigate all dependent append-only tables and token usage run_id SET NULL behavior before choosing a migration.

Start by reproducing the failure with a real PostgreSQL fixture containing a completed Agent run, then inspect all dependent append-only tables and token-usage run_id SET NULL behavior. Compare the related release-probe fixes in #3549 and #3551 without treating them as a solution. Done means authorized version-checked thread deletion or documented retention avoids HTTP 500 while preserving tampering protections, billing usage, tenant isolation, and organization teardown semantics.

Written by the indexing model from the issue text.