refactor(session): drop excerpt-cache narration

Author: TheGreatAxios

src/session/compactor-excerpt-once.test.ts

@@ -66,7 +66,7 @@ function liveResultText(turns: ConversationTurn[]): string {
     .join("\n");
 }
 
-describe("CL-10535 single excerpt per tail turn", () => {
+describe("excerpt once per tail turn", () => {
   test("compactor excerpts each tail turn exactly once with identical values", async () => {
     const dump = "z".repeat(12_000);
     const turns: ConversationTurn[] = [
@@ -80,8 +80,6 @@ describe("CL-10535 single excerpt per tail turn", () => {
 
     const before = excerptTailTurnCallCount();
     const result = await createPruningCompactor({
-      // Generous budget: every tail turn is a candidate, so a full pass
-      // must excerpt each turn exactly once.
       compactionShape: { tailBudgetTokens: 1_000_000 },
       summarize: async () => {
         throw new Error("must not invent a summary: nothing excluded");
@@ -90,13 +88,8 @@ describe("CL-10535 single excerpt per tail turn", () => {
 
     const calls = excerptTailTurnCallCount() - before;
 
-    // One excerptTailTurn call per tail turn — never the same turn twice
-    // (the closure estimate, turnCost, and emitted live copies share the
-    // first excerpt).
     expect(calls).toBe(turns.length);
 
-    // Values stay identical to the pre-fix behavior: the oversized outputs
-    // are still head+tail excerpted in the live tail.
     const live = liveResultText(result.output);
     expect(live.match(/\[tail-shortened \d+→/g)?.length ?? 0).toBe(3);
     expect(live).not.toContain(dump);

src/session/compactor.ts

@@ -647,10 +647,6 @@ function excerptTailText(
   };
 }
 
-// Test seam (CL-10535): total excerptTailTurn invocations since module load.
-// selectTail reuses the first excerpt per tail turn (closure estimates,
-// turnCost, and the emitted live copies share one cache), so a pass over N
-// tail turns adds exactly N to this counter.
 let excerptTailTurnCallCountValue = 0;
 export function excerptTailTurnCallCount(): number {
   return excerptTailTurnCallCountValue;
@@ -726,8 +722,6 @@ function lastAssistantToolCallIndex(
   return undefined;
 }
 
-// Token estimate over an already-excerpted live turn. Does not re-excerpt:
-// callers reuse the first excerptTailTurn result (CL-10535).
 function liveTurnTokens(live: ConversationTurn): number {
   let tokens = 0;
   for (const block of live.content) {
@@ -776,9 +770,6 @@ function selectTail(
   const budgetTokens = shape.tailBudgetTokens;
   const keepThinkingAt = lastAssistantToolCallIndex(turns);
 
-  // Excerpt results are cached per tail turn so each turn is excerpted at most
-  // once per pass: the closure token estimate, turnCost, and the emitted live
-  // copies all reuse the first excerpt (CL-10535).
   type ExcerptResult = ReturnType<typeof excerptTailTurn>;
   const excerptCache = new Map<number, ExcerptResult>();
   const excerptFor = (idx: number, turn: ConversationTurn): ExcerptResult => {