The `DecisionLedger.append` operation in the AI agent's memory system, specifically its non-atomic read-modify-write behavior leading to data loss under concurrency.

# DecisionLedger.append is a non-atomic read-modify-write

**Severity:** Medium
**File:** `packages/core/src/agent_memory/core/ledger.py:63`

`append()` reads the entire ledger file, appends one line, and writes the whole thing
back. Two concurrent `sleep` processes (or a `sleep` and a `decide` CLI call) can both
read the same content, each append their own decision, and the second write silently
drops the first decision. The file-level `store_lock` in `Store.write()` does not cover
the ledger, which lives under `dream_reports/`, not the truth directories.

```python
def append(self, decision: Decision) -> Decision:
self._path.parent.mkdir(parents=True, exist_ok=True)
head = self._path.read_text(encoding="utf-8") if self._path.exists() else HEADING + "\n\n"
self._path.write_text(head + decision.render() + "\n", encoding="utf-8")
return decision
```

### Why it matters
The ledger is truth, not cache (its own docstring says so). Losing a rejection verdict
means the same proposal re-appears on the next sleep and may get accepted by a reasoner
that sees it for the first time — the operator's explicit refusal is silently forgotten.