chore: remove recommendation section and internal architecture references

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Committed-By-Agent: claude

Author: kdhillon-stripe

docs/plans/2026-05-11-cache-sidecar-architecture.md

@@ -147,7 +147,7 @@ The current approach is time-based pruning with a fixed buffer. There is no way
 
 **Core framing:** For paying customers, denying valid usage is WORSE than allowing brief overage. Overage is billable and self-corrects on the next checkpoint. Denial loses revenue and erodes trust. Exception: free-tier and fraud-prevention contexts need strict enforcement (never permissive).
 
-**Background:** Metronome processes events across multiple partitions with variable lag. Out-of-order processing is real. External apps may send events the sidecar never sees.
+**Background:** Event processing has variable lag and out-of-order delivery is real. External apps may send events the sidecar never sees.
 
 ### Comparison
 
@@ -187,20 +187,13 @@ Query Metronome's usage API to confirm specific events appear in aggregated usag
 
 ### Watermark-based
 
-Metronome exposes `watermark_low` / `watermark_high` in balance responses. Prune events <= `watermark_low` (confirmed across all partitions). Events between low and high are ambiguous.
+Metronome exposes `watermark_low` / `watermark_high` in balance responses. Prune events <= `watermark_low` (confirmed processed). Events between low and high are ambiguous.
 
 - **Overage risk:** None in strict mode. Tiny in lenient mode (ambiguous-zone events).
 - **False denial risk:** None.
 - **Right choice:** Long-term production steady-state. Precise, minimal drift, no polling overhead.
 - **Deploy:** Blocked. Watermark metadata in balance responses is pending vendor support. Once available, days to integrate.
 
-### Recommendation
-
-- **Default (paying customers):** Time-based pruning + TTL floor in lenient mode. Brief overage is billable and self-corrects. Ship immediately.
-- **Strict (free-tier / fraud):** Same strategy, conservative tuning -- long TTL (10 min+), short buffer, strict mode. Deny rather than permit.
-- **Long-term:** Watermark-based. Precise, safe, minimal drift. Upgrade when available.
-- **High-value VIPs:** Event confirmation. Zero drift, provably correct. Reserve for accounts where API cost is justified.
-
 ---
 
 ## Pricing Strategy (as implemented)