Wires `executor.executions` into the Effect-native engine so every
`execute()` / `executeWithPause()` / `resume()` call writes an
`execution` row and its associated tool-call + interaction rows to
whichever `DBAdapter` backs the SDK.

Engine additions:
- `ExecutionTrigger` type + new `trigger?` option on `execute` and
  `executeWithPause`. Callers attribute runs ("cli", "http", "mcp",
  …); the kind + optional meta blob are persisted on the row.
- A stable `crypto.randomUUID()` execution id is minted at entry and
  reused as `PausedExecution.id`, so callers and the DB share the
  same identifier and counts line up across pause/resume.
- `makeRecordingInvoker` wraps the `SandboxToolInvoker` passed to the
  code executor; each `invoke` writes a tool-call row (running →
  completed|failed with duration). Storage errors are ignored so
  bookkeeping failures can never fail the tool call itself.
- `persistTerminalState` runs once on fiber success or failure and
  writes final status, result/error, logs, toolCallCount, completedAt.
- Pausable path: on elicitation, the execution transitions to
  `waiting_for_interaction` and a pending interaction row is created;
  `resume` resolves it (or cancels it if action === "cancel") before
  unblocking the fiber. A `toolCallCounters` map keeps the same Ref
  across pause/resume so the final count is accurate.
- Inline path: wraps the caller-supplied `onElicitation` so every
  inline elicitation gets the same pending → resolved bookkeeping.

Tests (`engine-persistence.test.ts`, 5 cases) cover:
- completed run + tool call rows
- error result → status=failed, errorText captured
- toolCallCount rolls up correctly
- trigger kind + meta persist on the row
- failed tool call records status=failed with errorText