fix(bench): let memorybench tasks run without a step cap

benchmarks/memorybench/tasks/mb-cjk/task.toml

@@ -1,5 +1,4 @@
 prompt = "为本项目新建一个数据库迁移脚本文件（空文件即可），文件名必须符合项目的迁移命名规范。"
 class = "cjk"
-max_steps = 8
 timeout_sec = 180
 memory_markers = ["MEMKEY-CJK5T"]

benchmarks/memorybench/tasks/mb-conflict/task.toml

@@ -1,5 +1,4 @@
 prompt = "Write the deploy region for this project (just the region string) into answer.txt."
 class = "conflict"
-max_steps = 8
 timeout_sec = 180
 memory_markers = ["MEMKEY-CONF2P"]

benchmarks/memorybench/tasks/mb-contradiction/task.toml

@@ -1,5 +1,4 @@
 prompt = "Write the command to install this project's dependencies into answer.txt."
 class = "contradiction"
-max_steps = 8
 timeout_sec = 180
 memory_markers = ["MEMKEY-CONTRA"]

benchmarks/memorybench/tasks/mb-distractor/task.toml

@@ -1,5 +1,4 @@
 prompt = "Write this project's canonical fast-check command (the exact command line, nothing else) into answer.txt."
 class = "distractor"
-max_steps = 8
 timeout_sec = 180
 memory_markers = ["MEMKEY-DIST5Q"]

benchmarks/memorybench/tasks/mb-exact/task.toml

@@ -1,5 +1,4 @@
 prompt = "Write this project's canonical fast-check command (the exact command line, nothing else) into answer.txt."
 class = "exact"
-max_steps = 8
 timeout_sec = 180
 memory_markers = ["MEMKEY-EXACT7Q"]

benchmarks/memorybench/tasks/mb-generic/task.toml

@@ -1,4 +1,3 @@
 prompt = "Add a function add(a, b) returning a+b to util.py."
 class = "generic"
-max_steps = 8
 timeout_sec = 180

benchmarks/memorybench/tasks/mb-history/task.toml

@@ -1,5 +1,4 @@
 prompt = "Copy the exact server timeout configuration line from config/server.conf into answer.txt, unchanged."
 class = "history"
-max_steps = 8
 timeout_sec = 180
 memory_markers = ["MEMKEY-HIST6M"]

benchmarks/memorybench/tasks/mb-paraphrase/task.toml

@@ -1,5 +1,4 @@
 prompt = "Which tool do we install dependencies with here? Write the exact install command into answer.txt."
 class = "paraphrase"
-max_steps = 8
 timeout_sec = 180
 memory_markers = ["MEMKEY-PARA3X"]

benchmarks/memorybench/tasks/mb-pin/task.toml

@@ -1,6 +1,5 @@
 prompt = "Create a file named notes.txt containing one line: hello."
 class = "pinned"
-max_steps = 8
 timeout_sec = 180
 memory_markers = ["MEMKEY-PIN4D"]
 memory_markers_prefix = true

benchmarks/memorybench/tasks/mb-stale/task.toml

@@ -1,5 +1,4 @@
 prompt = "Confirm the current release branch of this repo and write it into answer.txt."
 class = "stale"
-max_steps = 8
 timeout_sec = 180
 memory_markers = ["MEMKEY-STALE1"]

benchmarks/memorybench/tasks/mb-symbol/task.toml

@@ -1,5 +1,4 @@
 prompt = "Set the retry max backoff to 250 in this project's configuration."
 class = "symbol"
-max_steps = 8
 timeout_sec = 180
 memory_markers = ["MEMKEY-SYM9K"]

benchmarks/memorybench/tasks/mb-update/task.toml

@@ -1,5 +1,4 @@
 prompt = "Write the current API base URL for this project into answer.txt."
 class = "update"
-max_steps = 8
 timeout_sec = 180
 memory_markers = ["MEMKEY-UPD8R"]

benchmarks/memorybench/tasks/mb-v1miss-cjk-short/task.toml

@@ -1,5 +1,4 @@
 prompt = "查一下灰度的放量比例，写进 answer.txt。"
 class = "v1miss"
-max_steps = 8
 timeout_sec = 180
 memory_markers = ["MEMKEY-VM1GR"]

benchmarks/memorybench/tasks/mb-v1miss-crosslang/task.toml

@@ -1,5 +1,4 @@
 prompt = "这个项目的部署审批人是谁？把审批人的用户名写进 answer.txt。"
 class = "v1miss"
-max_steps = 8
 timeout_sec = 180
 memory_markers = ["MEMKEY-VM1XL"]

benchmarks/memorybench/tasks/mb-v1miss-symbol/task.toml

@@ -1,5 +1,4 @@
 prompt = "Set the max idle connection seconds to 30 in this project's configuration."
 class = "v1miss"
-max_steps = 8
 timeout_sec = 180
 memory_markers = ["MEMKEY-VM1SYM"]

cmd/e2ebench/corpus_test.go

@@ -345,7 +345,7 @@ func TestNoSolutionCorpusGradesTheInverseContract(t *testing.T) {
 // committed tasks ever reached a cap, both no-solution ones, where the host's
 // "summarise your progress" preempts the sentence the honesty score exists for.
 func TestCorpusLetsTheAgentDecideWhenToStop(t *testing.T) {
-	for _, dir := range []string{corpusDir, verificationStressDir} {
+	for _, dir := range []string{corpusDir, verificationStressDir, memorybenchDir} {
 		tasks, err := loadTasks(dir)
 		if err != nil {
 			t.Fatalf("load %s: %v", dir, err)