fix: preserve failed responses when retrying model calls

Keep failed attempts in the saved transcript and omit only trailing failures from Retry's temporary agent history. Use the matching temporary prefix length when persisting genuinely new output.

Preserve the error notice across plan/compaction markers, refuse Retry after later compaction, and avoid synthesizing results for tool calls inside failed generations while retaining Pi's native filtering.

Verified: 173 focused Retry/session/compaction-projection/practice-record tests, typecheck, changed-file lint, and diff --check. No app launch or real provider calls.

src/core/native-messages.ts

@@ -551,9 +551,10 @@ function resultIdsFrom(messages: readonly AgentMessage[]): Set<string> {
  * - markers flatten to a user message carrying their content (the
  *   canonical text single-sourced in core/contracts); plan blocks are
  *   UI-only and filtered.
- * - dangling tool calls (an aborted turn's unanswered questions) get the
- *   synthesized interrupted result — the provider protocol's hard
- *   invariant: a tool call must be followed by its result.
+ * - dangling calls from successfully generated assistant messages (for
+ *   example, an unanswered question) get a synthesized interrupted result.
+ *   Error/aborted generations do not: Pi drops those assistant messages,
+ *   so manufacturing their results would leave orphan results on the wire.
  *
  * Contract (pi-agent-core): never throws — an unknown shape passes
  * through untouched rather than killing the loop.
@@ -619,11 +620,12 @@ export function toLlmMessages(messages: readonly AgentMessage[]): Message[] {
 		// Strip the record id: the wire carries the native shape only.
 		const { id: _id, ...native } = m as LhMessage;
 		out.push(native as Message);
-		// Dangling-call synthesis: an assistant message whose toolCalls have
-		// no results in the remaining tail (abort mid-question) — the
-		// provider rejects unanswered calls outright.
+		// Repair unanswered calls from completed generations, not calls in
+		// failed generations. Pi filters error/aborted assistants itself;
+		// adding results for their unexecuted calls would leave orphans.
 		if (role === "assistant") {
 			const assistant = m as AssistantMessage;
+			if (assistant.stopReason === "error" || assistant.stopReason === "aborted") continue;
 			const calls = assistant.content.filter((b) => b.type === "toolCall");
 			if (calls.length === 0) continue;
 			const answered = resultIdsFrom(messages.slice(i + 1));