ctxscan — transcript token/reclaim scanner (per-agent, venus CC only)

/resource/llmmsg-srv/tools/ctxscan/{scan.py,reclaim.py,README.md}. One-off measurement tool (not a service): per-agent turns/cache_read/mean-max-ctx, compact-boundary reclaim-vs-regrowth, hourly mean-ctx. Findings 2026-08-22: raw transcript scans overcount tokens 1.52x (streamed usage-block duplicates, dedup on requestId - usage_daily/KPI is canonical); compaction reclaims fully, not a ratchet, the lever is per-agent ceiling not reclaim effectiveness. Coverage: venus CC sessions with a transcript that moved (24 of 47 roster rows) only - scripts/-ca/off-host absent from this AND monitor-context. Not a repo, not scheduled. Owner coder-llmmsgsrv-cc.