Collect PromptKit execution observations with logical attribution

docs/internal/llm-runtime.md

@@ -30,5 +30,9 @@ PromptKit's per-backend admission applies to both roles and every generation,
 retry, and repair. The runner still processes module instances in order,
 collects concurrent section work deterministically, and cancels remaining
 section work on a runtime failure. `llm.DiagnosticsWriter` remains the protected
-request/response artifact writer. Execution observations and report summaries
-are completed by the following roadmap work.
+request/response artifact writer. `llm.ObservationCollector` captures metadata
+events under a short mutex and freezes them after runner work joins. Its logical
+records include preparation failures; summary usage comes from terminal events
+once per call and retains knownness and completeness for each field. Runner wall
+and validator timings remain separate from provider attempts. The report and
+artifact mapping is completed by the following roadmap work.

docs/roadmap/implementation.md

@@ -429,7 +429,16 @@ or add placeholders merely to make this intermediate commit build.
 
 ## Stage 9: Collect Execution Observations Safely
 
-**Status:** Pending. **Depends on:** Stage 8.
+**Status:** Complete (2026-09-24). Added Audita-owned logical-call and execution
+event records with call IDs, diagnostic attribution, short synchronized capture,
+copying snapshots, and terminal-only usage aggregation. Preserved runner wall
+and per-validator timings. Tests cover preparation failure, undispatched
+admission, known zero versus unknown usage, partial failed repairs, no double
+counting, concurrent attribution, and the real role-client hook. `go test
+-race ./internal/framework/llm ./internal/framework/runner`, focused package
+tests, and `go build -o /tmp/audita-stage9-build ./cmd/audita` passed;
+`git diff --check` passed. Report/artifact serialization follows in Stage 10.
+**Depends on:** Stage 8.
 **Owns:** Audita-owned execution record types and collector in the LLM boundary,
 runner attribution, and concurrency/observation tests.