style(reactor): drop essay comments from task doom-loop exemption

src/provider/inference-dependencies.ts

@@ -110,10 +110,6 @@ export function createInferenceDependencies(): Promise<Dependencies> {
       .then((deps) => ({
         ...deps,
         fetch: withCodexContentTypeRepair(deps.fetch),
-        // Compose the first-party doom-loop exemptions into the single liveness
-        // hook the vendored reactor already consumes: still-pending polls and
-        // idempotent task bookkeeping both reset the repeat streak instead of
-        // counting toward the fail-run threshold.
         isPollOnlyPendingBatch: (calls, results) =>
           isPollOnlyPendingBatch(calls, results) ||
           isIdempotentTaskBatch(calls, results),

src/subagent/idempotent-task-exempt.ts

@@ -3,16 +3,9 @@ import type { ToolCall, ToolResult } from "@intx/types/runtime";
 
 /**
  * Doom-loop exemption for idempotent task bookkeeping. A batch is exempt only
- * when it is non-empty and every executed call is a task-bookkeeping tool
- * (`manage_tasks` — `isSameTool` folds the `todowrite` and `update_plan`
- * aliases onto that engine id).
- *
- * Repeating an identical task update is legitimate bookkeeping, not a runaway
- * loop: "finalize checklist" three times changes nothing and should not
- * abort an otherwise-finished run under the `fail-run` policy. This resets
- * the doom-loop streak (same mechanism as the poll exemption) so consecutive
- * task-only batches never accumulate toward the threshold, while any real
- * tool in the batch counts normally and always preserves the guard.
+ * when it is non-empty and every call is a task-bookkeeping tool
+ * (`manage_tasks` — `isSameTool` folds `todowrite` and `update_plan`).
+ * Mixed batches return false so the guard counts them normally.
  */
 export function isIdempotentTaskBatch(
   calls: readonly ToolCall[],