docs(lessons): extend measure-before-optimize for PR/comment claims

lessons/workflow/measure-before-optimize.md

@@ -7,6 +7,11 @@ match:
   - this looks complex so it must be slow
   - pytest --profile
   - premature optimization
+  - claiming latency win without measurement
+  - PR description promises speedup
+  - expected speedup not verified
+  - context size is the bottleneck
+  - skipping context will speed this up
 status: active
 ---
 
@@ -24,6 +29,8 @@ Observable signals indicating need for measurement:
 - Assuming slowness based on complexity without measurement
 - Building caching solutions before confirming problems exist
 - Unable to articulate specific performance metrics that need improvement
+- Writing "should drop to Xs" in a PR description or issue comment without a baseline number
+- Pointing at one component (prompt size, import time, N+1) as "the" bottleneck without having compared it to other candidates
 
 ## Pattern
 Measure first, then decide based on data:
@@ -49,6 +56,21 @@ pytest tests/test_lessons*.py --profile --durations=10
 # No profiling data to support the need
 ```
 
+**Anti-pattern**: Latency-win claims in PR descriptions / issue comments
+```text
+# smell: PR body says "simple lookups should drop to ~5-10s"
+# but no before/after timing is shown, and the target number
+# is anchored on one component (prompt size) rather than on
+# the actual cost model (turns × tok/s + startup + tool-use).
+
+# Correct shape:
+# - show baseline (e.g. "currently p50 = 42s over N calls")
+# - explain cost model and which component dominates
+# - project delta from the dominant component, not from the
+#   smallest one
+# - after landing, post measured delta — not re-projected one
+```
+
 ## Outcome
 Following this pattern leads to:
 - **Evidence-based decisions**: Optimize what actually matters