v5.11.18 — frozen savings baseline

Frozen per-session savings baseline

The "old way / session" figure used to float every run because it priced the current 30-day window volume. It is now a frozen typical-session anchor: it depends only on captured baseline inputs + prices, so it stays put run to run. "Now / session" moves only with efficiency (model mix + cache reuse); monthly = per-session cut × session count.

Changes

• Stable baseline across all engines (Python = Claude Code + Codex; OpenClaw; OpenCode). Platform parity restored.
• Fixed the dashboard tokens/session mislabel (was the monthly total, now the true per-session footprint).
• Disambiguated the two "Context Quality" cards: Savings tab = "30-day average", Health tab = "latest session".
• Hardening (torture room, 0 critical / 20-of-20 chaos pass): ArithmeticError-safe catch, now-cost guard, net_negative field population, dead-kwarg removal.

Tests

• 11 Python baseline-stability tests + 4 OpenCode bun test cases proving the frozen-baseline invariant; OpenClaw verified via Node check.