fix: distinguish dependency readiness from code verification

LESSONS.md

@@ -13,6 +13,13 @@ Dated, short lessons from real FlyBot work. Agents read this **before coding**.
 
 ## Open lessons
 
+### 2026-11-04 — An offline dependency miss is not an assertion failure (Seen 1×)
+
+Prepare and observe each executor separately, then verify the exact cleaned candidate.
+Keep non-success classified and old evidence immutable. A killed docker client is
+not proof its preparation command exited; retain the original occupancy fence.
+
+
 ### 2026-11-04 — Parsing success is not effective permission proof (Seen 1×)
 
 CLI 0.166 accepts unknown filesystem alias keys. Use supported `:tmpdir`, inspect

PROJECT.md

@@ -16,6 +16,8 @@ branch. #2336 establishes owned reservation rollback and a durable pre-spawn fen
 unknown execution, workspace/context and current authority remain protected.
 #2337 validates the actual SDK-resolved CLI and role profile before inference;
 configuration readiness is separate from model/test/exit evidence.
+#2334 separates executor dependency readiness from host test outcomes and preserves
+same-candidate retry evidence; remote unknown exit retains preparation occupancy.
 No live task/configuration/credential mutation or deployment is implied.
 
 ## Execution recovery — 2026-11-04

ROADMAP.md

@@ -6,7 +6,8 @@ Implement #2336 reservation cleanup first; then #2337 CLI configuration, #2334
 environment attribution, #2346 context migration, #2338 receipts, #2347 recovery
 routing, #2331 controlled merge and #2342 operational-agent collaboration. Separate
 PRs retain existing UI/control-plane, authority and exact exit/evidence contracts.
-#2336 is merged; #2337 adds offline actual-CLI readiness and scratch protection.
+#2336/#2337 are merged. #2334 adds executor dependency receipts and classified
+verification; actual Docker and paid inference remain distinct acceptance evidence.
 
 ## Recovery delivery — 2026-11-04
 

changelog.d/2334.fixed.md

@@ -0,0 +1 @@
+Record executor-scoped dependency readiness and classify verification outcomes; preserve same-candidate retry history and retain remote preparation capacity when exit is unknown.