benchmarks: disk readings are persisted footprints (FalkorDB relaxed SAVE, no lifecycle disk, no negative growth) (#349) - #350

Part 1 of #349 (the harness). Part 2, the publish gate and the shared-memory disclosure, follows in its own PR.

• A relaxed FalkorDB cell runs redis-cli SAVE in the server before its final disk reading (pre_disk_cmd on both FalkorDB backends, the result on the row as server_pre_disk). That class persists only when a Redis snapshot rule fires, so a cell that ended first read 0.0 MB beside its neighbours' gigabytes (l3d small, l2 sf1). The strict class is untouched: its append-only file already holds every write.
• The lifecycle lane records no disk_data_mb: it removes each database after measuring it, so its reading was the empty container at every size.
• A growth below zero records no disk_data_mb, with disk_note saying why (QuestDB restart ts100 read down to -95.9 MB because its empty-start reading holds preallocated files).
• PROTOCOL.md carries both rules.

What it does to existing rows: nothing is edited. The two FalkorDB relaxed cells (l3d small, l2 sf1 OLTP) are re-run on mini after this merges, as a repair stage after qR18; the lifecycle and restart readings are not on the page.

Checked: the ten test files that import runner (377 passed), uvx pre-commit run on the changed files.

🤖 Generated with Claude Code