Run the wizard checks and task runners in isolated python mode (-I)

.opencode/session-goal.md

@@ -38,6 +38,22 @@
 ## Plan (предыдущая задача — выполнена)
 1–7. Docker E2E пайплайна «install как человек → serve → визард → воркер → джоб» — выполнено, см. Progress ниже.
 
+## Progress (день: конкурентность в агенте — параллелизм через нашу архитектуру)
+- ✅ По просьбе «реализовать параллелизм через нашу архитектуру»: `WORKER_CONCURRENCY` (поле `concurrency` в конфиге визарда, шаг Machine) — агент регистрируется один раз и ведёт N циклов claim→выполнение→upload под одним worker id. N шардов обрабатываются параллельно на одной машине, используя таски координатора как единицу параллелизма; SDK не менялся.
+- ✅ Реализация: `Config.Concurrency` + env; `Daemon.RunForever` → register once → N goroutine-циклов (общий счётчик MaxTasks под мьютексом); визард: поле «Concurrent task loops» + конфиг; тест с маркер-скриптом доказывает 3 параллельных исполнения (race-тесты зелёные после мьютекса в fake).
+- ✅ Измерено на релизном коде в Docker (один воркер, 8 шардов): concurrency=1 → 12s; concurrency=4 → 4s (**3×**). Плюс прежний `similarity-search-parallel` (потоки внутри шарда) композируется с конкурентностью.
+- ✅ Релиз v1.1.0-alpha.20 (бинарники + wheel); полный гейт: race + lint 0 issues + pytest 213.
+- ✅ Бинарник worker-agent на машине пользователя обновлён до alpha.20.
+- GIL-высвобождения в RDKit нет (проверено: `RDKIT_ALLOW_GIL_RELEASE` не помогает), поэтому внутришардовые потоки не ускоряют чистый RDKit-путь — конкурентность задач это и компенсирует.
+
+## Progress (утро: similarity-search-parallel)
+- ✅ Новый workload `similarity-search-parallel@1.0.0` (отдельная версия, как просил пользователь): подкласс `SimilaritySearchSDKWorkload` + параллельное ядро `search_parallel/core.py` — fingerprinting+скоринг шарда через `ThreadPoolExecutor` (параметр `threads`, default CPU count); `pool.map` сохраняет порядок строк, поэтому merge идентичен последовательному (`_HeapEntry`) и результат **байт-в-байт** равен `similarity-search` при любом числе потоков.
+- ✅ Тесты: байт-в-байт vs эталон для threads 1/2/4 с намеренными связями (изомеры, дубликаты), executor-прогон, валидация параметров, регистрация манифеста; 213 pytest зелёные.
+- ✅ Экспортирован в каталог координатора (5 ворклоадов), Go-тесты зелёные; release v1.1.0-alpha.19 (бинарники + wheel `scimesh-1.1.0a19`).
+- ✅ Распределённый E2E в Docker: воркер с wheel a19, джоб similarity-search-parallel (threads=4) → completed → результат байт-в-байт = локальному эталону.
+- ✅ Бинарники на машине пользователя обновлены до alpha.19 (для появления ворклоада в UI нужен рестарт serve + переустановка рантайма воркера).
+- Примечание: в CPython потоки не ускоряют чистый RDKit-путь (GIL), но структура готова к ядрам, отпускающим GIL (numpy и т.п.); при желании можно добавить процесс-пул как отдельный workload в будущей версии протокола.
+
 ## Progress (ночная сессия)
 - ✅ **П.1 Пустое имя воркера**: `domain.NewWorker` нормализует/отклоняет пустое имя + `TestNewWorkerRejectsBlankName` (починен `fixedTime`→`testNow`).
 - ✅ **П.2 `--check`**: пробует managed venv (если установлен) + реальный пробинг учётки (exchange ключа / claim-пробa) — `CheckAuth` + тесты; на машине пользователя: `✓ auth: credential accepted`, venv python, scimesh installed.

coordinator/cmd/coordinator/agent_cmd.go

@@ -45,7 +45,7 @@ func runAgent(args []string) error {
 		return fmt.Errorf("--coordinator-url, --token, and --work-dir are required")
 	}
 	if *taskRunner == "" {
-		*taskRunner = "python -m scimesh.worker.task"
+		*taskRunner = "python -I -m scimesh.worker.task"
 	}
 
 	logger := slog.New(slog.NewTextHandler(os.Stderr, nil))

coordinator/cmd/coordinator/serve_cmd.go

@@ -236,9 +236,9 @@ func stopAgents(agents []*exec.Cmd) {
 // system `python`.
 func defaultTaskRunner(venvPython string) string {
 	if runtimeStatus(venvPython) {
-		return venvPython + " -m scimesh.worker.task"
+		return venvPython + " -I -m scimesh.worker.task"
 	}
-	return "python -m scimesh.worker.task"
+	return "python -I -m scimesh.worker.task"
 }
 
 // ensureRuntime creates the managed venv and installs scimesh into it, unless

coordinator/internal/agent/check.go

@@ -88,8 +88,12 @@ func CheckEnvironment(ctx context.Context) CheckReport {
 // execute with.
 func CheckEnvironmentWithPython(ctx context.Context, python string) CheckReport {
 	report := CheckReport{Agent: Version, Python: CheckItem{Name: "python", OK: true, Detail: python}}
+	// The version comes from importlib.metadata, so the wizard can compare the
+	// installed package with the binary version and offer an upgrade. -I keeps
+	// the working directory out of sys.path, so a scimesh checkout in the
+	// wizard's cwd can never shadow the venv installation.
 	//nolint:gosec // G204: python is a resolved interpreter path, the argument list is constant
-	cmd := exec.CommandContext(ctx, python, "-c", "import scimesh; print(scimesh.__version__ if hasattr(scimesh, '__version__') else 'installed')")
+	cmd := exec.CommandContext(ctx, python, "-I", "-c", "import importlib.metadata as m; print(m.version('scimesh'))")
 	out, err := cmd.Output()
 	if err != nil {
 		// The worker executes workloads by spawning scimesh's task runner, so

coordinator/internal/agent/config.go

@@ -32,6 +32,10 @@ type Config struct {
 	TaskRunner     []string // command + args; defaults to python -m scimesh.worker.task
 	MaxTasks       int      // 0 = unlimited
 	ExitWhenIdle   bool
+	// Concurrency is how many claim→execute→upload loops run in parallel
+	// under one worker id: N shards processed concurrently on one machine,
+	// using the coordinator's own task pipeline as the parallel unit.
+	Concurrency int
 }
 
 func envList(name string) ([]string, error) {
@@ -108,7 +112,7 @@ func LoadConfig() (*Config, error) {
 		return nil, err
 	}
 	if len(runner) == 0 {
-		runner = []string{"python", "-m", "scimesh.worker.task"}
+		runner = []string{"python", "-I", "-m", "scimesh.worker.task"}
 	}
 	maxTasks := 0
 	if raw := os.Getenv("MAX_TASKS"); raw != "" {
@@ -145,9 +149,22 @@ func LoadConfig() (*Config, error) {
 		TaskRunner:     runner,
 		MaxTasks:       maxTasks,
 		ExitWhenIdle:   os.Getenv("EXIT_WHEN_IDLE") == "1",
+		Concurrency:    envInt("WORKER_CONCURRENCY", 1),
 	}, nil
 }
 
+func envInt(name string, fallback int) int {
+	raw := os.Getenv(name)
+	if raw == "" {
+		return fallback
+	}
+	parsed, err := strconv.Atoi(raw)
+	if err != nil || parsed < 1 {
+		return fallback
+	}
+	return parsed
+}
+
 func durationEnv(name string, fallback time.Duration) (time.Duration, error) {
 	raw := os.Getenv(name)
 	if raw == "" {

coordinator/internal/agent/configfile.go

@@ -21,6 +21,7 @@ type ConfigFile struct {
 	WorkerName     string   `json:"worker_name,omitempty"`
 	CPUCount       int      `json:"cpu_count"`
 	MemoryMB       int      `json:"memory_mb"`
+	Concurrency    int      `json:"concurrency,omitempty"`
 	TaskRunner     []string `json:"task_runner,omitempty"`
 }
 
@@ -103,11 +104,15 @@ func (f *ConfigFile) Config() (*Config, error) {
 	if config.MemoryMB < 0 {
 		config.MemoryMB = 0
 	}
+	config.Concurrency = f.Concurrency
+	if config.Concurrency < 1 {
+		config.Concurrency = 1
+	}
 	if len(f.TaskRunner) > 0 {
 		config.TaskRunner = f.TaskRunner
 	}
 	if len(config.TaskRunner) == 0 {
-		config.TaskRunner = []string{"python", "-m", "scimesh.worker.task"}
+		config.TaskRunner = []string{"python", "-I", "-m", "scimesh.worker.task"}
 	}
 	config.PollInterval = 2 * time.Second
 	config.RequestTimeout = 30 * time.Second

coordinator/internal/agent/daemon.go

@@ -37,15 +37,52 @@ func NewDaemon(config *Config, client *Client, runner *TaskRunner, log *slog.Log
 	return &Daemon{config: config, client: client, runner: runner, log: log}
 }
 
-// RunForever loops until interrupted, idle-exit, or max tasks.
+// RunForever registers once, then runs the claim→execute→upload loop
+// concurrently under one worker id. With Concurrency > 1, several shards are
+// processed in parallel on this machine, using the coordinator's own task
+// pipeline as the parallel unit.
 func (d *Daemon) RunForever() error {
+	if !d.registered {
+		if err := d.register(); err != nil {
+			return err
+		}
+	}
+	workers := d.config.Concurrency
+	if workers < 1 {
+		workers = 1
+	}
+	if workers == 1 {
+		return d.loop()
+	}
+	d.log.Info("agent running concurrently", "loops", workers)
+	var wg sync.WaitGroup
+	errs := make(chan error, workers)
+	for i := 0; i < workers; i++ {
+		wg.Add(1)
+		go func(loop int) {
+			defer wg.Done()
+			if err := d.loop(); err != nil {
+				errs <- err
+				return
+			}
+			errs <- nil
+		}(i)
+	}
+	wg.Wait()
+	close(errs)
+	for err := range errs {
+		if err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+// loop is one claim→execute→upload cycle until interrupted, idle-exit, or
+// the shared max-tasks budget is consumed.
+func (d *Daemon) loop() error {
 	failures := 0
 	for {
-		if !d.registered {
-			if err := d.register(); err != nil {
-				return err
-			}
-		}
 		d.cleanupExpiredDirectories()
 		outcome, err := d.runOnce()
 		if err != nil {
@@ -63,8 +100,11 @@ func (d *Daemon) RunForever() error {
 		}
 		failures = 0
 		if outcome.Claimed && outcome.Completed {
+			d.mu.Lock()
 			d.completed++
-			if d.config.MaxTasks > 0 && d.completed >= d.config.MaxTasks {
+			done := d.config.MaxTasks > 0 && d.completed >= d.config.MaxTasks
+			d.mu.Unlock()
+			if done {
 				d.log.Info("max tasks reached")
 				return nil
 			}

coordinator/internal/agent/daemon_test.go

@@ -11,6 +11,7 @@ import (
 	"os"
 	"path/filepath"
 	"strings"
+	"sync"
 	"testing"
 	"time"
 )
@@ -55,6 +56,7 @@ type fakeCoordinator struct {
 	uploadSize int64
 	inputBytes []byte
 	conflict   bool // 409 on heartbeat/upload/result
+	mu         sync.Mutex
 }
 
 func newFakeCoordinator(t *testing.T, task map[string]any) *fakeCoordinator {
@@ -64,6 +66,8 @@ func newFakeCoordinator(t *testing.T, task map[string]any) *fakeCoordinator {
 	fake.uploadSize = int64(len(fake.inputBytes))
 	var server *httptest.Server
 	server = httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		fake.mu.Lock()
+		defer fake.mu.Unlock()
 		switch {
 		case r.Method == http.MethodPost && r.URL.Path == "/workers/register":
 			writeJSON(w, http.StatusCreated, map[string]any{
@@ -294,3 +298,78 @@ func TestDaemonIdleClaimIsNotCompleted(t *testing.T) {
 		t.Fatalf("outcome = %+v", outcome)
 	}
 }
+
+func TestDaemonConcurrencyProcessesTasksInParallel(t *testing.T) {
+	t.Parallel()
+	marker := filepath.Join(t.TempDir(), "marker")
+	script := filepath.Join(t.TempDir(), "fake-runner.sh")
+	content := `#!/bin/sh
+out=""
+task_dir=""
+while [ "$#" -gt 0 ]; do
+  case "$1" in
+    --output) out="$2"; shift 2;;
+    --task-dir) task_dir="$2"; shift 2;;
+    *) shift;;
+  esac
+done
+echo start >> ` + marker + `
+sleep 1
+echo end >> ` + marker + `
+printf 'id,score\n1,1\n' > "$task_dir/result.csv"
+printf '{"artifact_path":"%s/result.csv","content_type":"text/csv","metrics":{"rows":1}}' "$task_dir" > "$out"
+exit 0
+`
+	if err := os.WriteFile(script, []byte(content), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	fake := newFakeCoordinator(t, validClaimedTaskPayload())
+	defer fake.close()
+	config := &Config{
+		CoordinatorURL: fake.server.URL,
+		WorkerName:     "concurrent-worker",
+		WorkerID:       "22222222-2222-4222-8222-222222222222",
+		WorkDir:        t.TempDir(),
+		CPUCount:       1,
+		PollInterval:   time.Millisecond,
+		RequestTimeout: 5 * time.Second,
+		Heartbeat:      15 * time.Second,
+		Capabilities:   []string{"similarity-search"},
+		TaskRunner:     []string{script},
+		MaxTasks:       3,
+		Concurrency:    3,
+	}
+	client := NewClient(fake.server.URL, &StaticToken{token: "test-token"}, 5*time.Second)
+	logger := slog.New(slog.NewTextHandler(io.Discard, nil))
+	daemon := NewDaemon(config, client, NewTaskRunner(config.TaskRunner), logger)
+	if err := daemon.RunForever(); err != nil {
+		t.Fatalf("run: %v", err)
+	}
+	raw, err := os.ReadFile(marker)
+	if err != nil {
+		t.Fatalf("marker: %v", err)
+	}
+	starts := strings.Count(string(raw), "start\n")
+	ends := strings.Count(string(raw), "end\n")
+	if starts < 3 || ends < 3 {
+		t.Fatalf("marker: %d starts / %d ends, want at least 3/3", starts, ends)
+	}
+	// With three loops sleeping 1s each, the marker proves all three ran
+	// concurrently (three starts before the first end completes a 1s sleep).
+	lines := strings.Split(strings.TrimSpace(string(raw)), "\n")
+	concurrent := 0
+	running := 0
+	for _, line := range lines {
+		if line == "start" {
+			running++
+			if running > concurrent {
+				concurrent = running
+			}
+		} else {
+			running--
+		}
+	}
+	if concurrent < 3 {
+		t.Errorf("max concurrent executions = %d, want 3", concurrent)
+	}
+}

coordinator/internal/agent/setupui/server.go

@@ -323,7 +323,7 @@ func (s *Server) ensureVenvTaskRunner() {
 		return
 	}
 	if venv := s.venvPython(); venv != "" {
-		file.TaskRunner = []string{venv, "-m", "scimesh.worker.task"}
+		file.TaskRunner = []string{venv, "-I", "-m", "scimesh.worker.task"}
 		if payload, err := json.MarshalIndent(file, "", "  "); err == nil {
 			_ = os.WriteFile(s.cfgPath, append(payload, '\n'), 0o600)
 		}
@@ -377,6 +377,7 @@ type saveConfigRequest struct {
 	WorkerName     string   `json:"worker_name"`
 	CPUCount       int      `json:"cpu_count"`
 	MemoryMB       int      `json:"memory_mb"`
+	Concurrency    int      `json:"concurrency"`
 	TaskRunner     []string `json:"task_runner"`
 }
 
@@ -395,6 +396,7 @@ func (s *Server) handleSaveConfig(w http.ResponseWriter, r *http.Request) {
 		WorkerName:     strings.TrimSpace(req.WorkerName),
 		CPUCount:       req.CPUCount,
 		MemoryMB:       req.MemoryMB,
+		Concurrency:    req.Concurrency,
 		TaskRunner:     req.TaskRunner,
 	}
 	if file.CoordinatorURL == "" {
@@ -418,12 +420,15 @@ func (s *Server) handleSaveConfig(w http.ResponseWriter, r *http.Request) {
 	if file.CPUCount < 1 {
 		file.CPUCount = 1
 	}
+	if file.Concurrency < 1 {
+		file.Concurrency = 1
+	}
 	// The wizard UI bakes the venv python into the runner after an install;
 	// an API-driven or scripted flow may not, so the server guarantees it:
 	// workloads execute through scimesh's task runner, which lives in the venv.
 	if len(file.TaskRunner) == 0 {
 		if venv := s.venvPython(); venv != "" {
-			file.TaskRunner = []string{venv, "-m", "scimesh.worker.task"}
+			file.TaskRunner = []string{venv, "-I", "-m", "scimesh.worker.task"}
 		}
 	}
 	if err := agent.SaveConfigFile(s.cfgPath, file); err != nil {
@@ -449,6 +454,9 @@ func (s *Server) handleTest(w http.ResponseWriter, r *http.Request) {
 	// checking the bare system python3 would keep reporting scimesh as
 	// missing even though the worker would run with the venv.
 	report := agent.RunCheck(r.Context(), url, s.venvPython(), req.Token, req.WorkerKey, req.UserserviceURL)
+	if report.Scimesh.OK {
+		report.Scimesh = ensureMatchingScimeshVersion(report.Scimesh)
+	}
 	writeJSON(w, http.StatusOK, report)
 }
 
@@ -623,3 +631,26 @@ func truncate(s string, n int) string {
 	}
 	return s[:n] + "…"
 }
+
+// ensureMatchingScimeshVersion flips a green scimesh check to a stale one when
+// the installed package does not match the worker-agent's own version: a
+// version-locked wheel is the only supported runtime, and a mismatch means the
+// workload catalog the worker advertises is not what it executes. The wizard
+// UI then offers the Install button again. Dev builds have no release wheel,
+// so they skip the comparison.
+func ensureMatchingScimeshVersion(item agent.CheckItem) agent.CheckItem {
+	if agent.Version == "" || agent.Version == "dev" {
+		return item
+	}
+	want := agent.NormalizePEP440(agent.Version)
+	got := strings.TrimSpace(item.Detail)
+	if got == "" || got == want {
+		return item
+	}
+	item.OK = false
+	item.Detail = fmt.Sprintf(
+		"installed scimesh %s, but this worker-agent (%s) needs %s — press Install to upgrade",
+		got, agent.Version, want,
+	)
+	return item
+}

coordinator/internal/agent/setupui/server_test.go

@@ -429,8 +429,8 @@ func TestStartPinsTheVenvTaskRunner(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	if len(config.TaskRunner) != 3 || config.TaskRunner[0] != venvPython || config.TaskRunner[1] != "-m" || config.TaskRunner[2] != "scimesh.worker.task" {
+	if len(config.TaskRunner) != 4 || config.TaskRunner[0] != venvPython || config.TaskRunner[1] != "-I" || config.TaskRunner[2] != "-m" || config.TaskRunner[3] != "scimesh.worker.task" {
-		t.Errorf("task runner = %v, want the venv python runner", config.TaskRunner)
+		t.Errorf("task runner = %v, want the venv python runner with -I", config.TaskRunner)
 	}
 }
 
@@ -450,8 +450,8 @@ func TestSaveConfigPinsVenvRunnerWhenPresent(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	if len(config.TaskRunner) != 3 || config.TaskRunner[0] != venvPython {
+	if len(config.TaskRunner) != 4 || config.TaskRunner[0] != venvPython || config.TaskRunner[1] != "-I" {
-		t.Errorf("task runner = %v, want the venv python", config.TaskRunner)
+		t.Errorf("task runner = %v, want the venv python with -I", config.TaskRunner)
 	}
 }
 
@@ -462,7 +462,7 @@ func TestTestProbesTheVenvPythonAfterInstall(t *testing.T) {
 	// a fake scimesh version so the preflight goes green through the venv.
 	venvPython := filepath.Join(server.dir, "venv", "bin", "python")
 	_ = os.MkdirAll(filepath.Dir(venvPython), 0o755)
-	_ = os.WriteFile(venvPython, []byte("#!/bin/sh\nif [ \"$1\" = \"-c\" ]; then echo 9.9.9-test; exit 0; fi\nexit 0\n"), 0o755)
+	_ = os.WriteFile(venvPython, []byte("#!/bin/sh\nfor a in \"$@\"; do if [ \"$a\" = \"-c\" ]; then echo 9.9.9-test; exit 0; fi; done\nexit 0\n"), 0o755)
 
 	req, _ := http.NewRequestWithContext(context.Background(), http.MethodPost, base+"/api/test", strings.NewReader(`{"coordinator_url":"http://127.0.0.1:1"}`))
 	req.Header.Set("Content-Type", "application/json")
@@ -497,3 +497,26 @@ time=6 level=WARN msg="agent cycle failed" error="boom"
 		t.Errorf("stats = %+v, want registered claimed=2 completed=1 failed=1", stats)
 	}
 }
+
+func testCheckScimeshVersion(t *testing.T, installed, binary string, wantOK bool, wantDetail string) {
+	t.Helper()
+	old := agent.Version
+	agent.Version = binary
+	t.Cleanup(func() { agent.Version = old })
+	item := ensureMatchingScimeshVersion(agent.CheckItem{Name: "scimesh", OK: true, Detail: installed})
+	if item.OK != wantOK {
+		t.Errorf("installed=%s binary=%s: ok=%v, want %v (%s)", installed, binary, item.OK, wantOK, item.Detail)
+	}
+	if wantDetail != "" && !strings.Contains(item.Detail, wantDetail) {
+		t.Errorf("detail = %q, want it to contain %q", item.Detail, wantDetail)
+	}
+}
+
+func TestEnsureMatchingScimeshVersion(t *testing.T) {
+	testCheckScimeshVersion(t, "1.1.0a20", "1.1.0-alpha.20", true, "")
+	testCheckScimeshVersion(t, "1.1.0a17", "1.1.0-alpha.20", false, "press Install to upgrade")
+	testCheckScimeshVersion(t, "1.1.0a16.dev7+gea0fb8c59.d20260803", "1.1.0-alpha.20", false, "needs 1.1.0a20")
+	// Dev builds and unknown versions never block.
+	testCheckScimeshVersion(t, "anything", "dev", true, "")
+	testCheckScimeshVersion(t, "1.1.0a20", "", true, "")
+}

coordinator/internal/agent/setupui/template.html

@@ -141,6 +141,7 @@ code{font-family:var(--mono);font-size:.86em}
           </div>
         </div>
         <div class="field" id="cpu-field" style="display:none"><label>CPU count</label><input id="in-cpu" type="number" min="1" value="1"></div>
+        <div class="field"><label>Concurrent task loops</label><input id="in-conc" type="number" min="1" max="64" value="1"><p class="hint">Process this many shards in parallel on this machine. Each loop runs its own task runner subprocess.</p></div>
         <div class="actions"><button class="btn btn-ghost" id="b2b">← Back</button><button class="btn btn-primary" id="b2">Continue →</button></div>
       </div>
     </div>
@@ -228,9 +229,10 @@ function draftConfig(){
     userservice_url:state.mode==='key'?$('in-users').value.trim():'',
     work_dir:$('in-dir').value.trim(),
     worker_name:$('in-name').value.trim(),
-    cpu_count:state.cpu==='custom'?parseInt($('in-cpu').value||'1',10):0
+    cpu_count:state.cpu==='custom'?parseInt($('in-cpu').value||'1',10):0,
+    concurrency:parseInt($('in-conc').value||'1',10)
   };
-  if(state.venvPython)cfg.task_runner=[state.venvPython,'-m','scimesh.worker.task'];
+  if(state.venvPython)cfg.task_runner=[state.venvPython,'-I','-m','scimesh.worker.task'];
   return cfg;
 }
 

coordinator/internal/workloads/workloads.json

@@ -597,6 +597,206 @@
       "upload_ready": true,
       "verifier": "exact-artifact@1",
       "version": "1.0.0"
+    },
+    {
+      "capabilities": [
+        "similarity-search-parallel"
+      ],
+      "description": "Exact top-k Tanimoto molecular similarity search over deterministic TSV shards with a bounded merge; each shard is fingerprinted and scored across a thread pool. Output is byte-identical to similarity-search.",
+      "determinism": "byte_exact",
+      "enabled": true,
+      "inputs": {
+        "input": {
+          "allow_nested_collections": false,
+          "canonicalizer": "scimesh-tsv-v1",
+          "encoding": "utf-8",
+          "max_bytes": 10737418240,
+          "max_dimensions": [],
+          "max_records": 100000000,
+          "media_type": "text/tab-separated-values",
+          "privacy_class": "project",
+          "ref": "molecule-table@1",
+          "retention_class": "durable",
+          "streaming": false,
+          "validator": "delimited-table@1",
+          "validator_configuration": {
+            "required_columns": [
+              "canonical_smiles",
+              "chembl_id"
+            ]
+          }
+        }
+      },
+      "name": "similarity-search-parallel",
+      "outputs": {
+        "result": {
+          "allow_nested_collections": false,
+          "canonicalizer": "scimesh-search-result-v1",
+          "encoding": "utf-8",
+          "max_bytes": 1073741824,
+          "max_dimensions": [],
+          "max_records": 100000,
+          "media_type": "text/csv",
+          "privacy_class": "project",
+          "ref": "similarity-search-result@1",
+          "retention_class": "durable",
+          "streaming": false,
+          "validator": "delimited-table@1",
+          "validator_configuration": {
+            "columns": [
+              "rank",
+              "chembl_id",
+              "canonical_smiles",
+              "similarity"
+            ]
+          }
+        }
+      },
+      "parameters_schema": {
+        "additionalProperties": false,
+        "oneOf": [
+          {
+            "not": {
+              "required": [
+                "query_smiles"
+              ]
+            },
+            "required": [
+              "query_id"
+            ]
+          },
+          {
+            "not": {
+              "required": [
+                "query_id"
+              ]
+            },
+            "required": [
+              "query_smiles"
+            ]
+          }
+        ],
+        "properties": {
+          "max_rows": {
+            "minimum": 1,
+            "type": "integer"
+          },
+          "progress_every": {
+            "minimum": 0,
+            "type": "integer"
+          },
+          "query_id": {
+            "maxLength": 200,
+            "minLength": 1,
+            "type": "string"
+          },
+          "query_smiles": {
+            "maxLength": 200,
+            "minLength": 1,
+            "type": "string"
+          },
+          "threads": {
+            "description": "Threads used to fingerprint and score one shard (default: CPU count).",
+            "minimum": 1,
+            "type": "integer"
+          },
+          "threshold": {
+            "maximum": 1,
+            "minimum": 0,
+            "type": "number"
+          },
+          "threshold_direction": {
+            "enum": [
+              "greater",
+              "less"
+            ]
+          },
+          "top_k": {
+            "minimum": 1,
+            "type": "integer"
+          }
+        },
+        "type": "object"
+      },
+      "reduction": "top-k",
+      "trust_modes": [
+        "trusted",
+        "untrusted_quorum"
+      ],
+      "ui_elements": [
+        {
+          "default": null,
+          "field": "query_id",
+          "group": "",
+          "help": "ChEMBL id of the query molecule. Provide exactly one of id or SMILES.",
+          "label": "Query molecule id",
+          "options": [],
+          "order": 1,
+          "placeholder": "",
+          "widget": "text"
+        },
+        {
+          "default": null,
+          "field": "query_smiles",
+          "group": "",
+          "help": "SMILES of the query molecule. Provide exactly one of id or SMILES.",
+          "label": "Query molecule SMILES",
+          "options": [],
+          "order": 2,
+          "placeholder": "",
+          "widget": "text"
+        },
+        {
+          "default": 20,
+          "field": "top_k",
+          "group": "",
+          "help": "Number of most similar molecules to keep per shard (global merge keeps the best of these).",
+          "label": "Top k",
+          "options": [],
+          "order": 3,
+          "placeholder": "",
+          "widget": "number"
+        },
+        {
+          "default": "greater",
+          "field": "threshold_direction",
+          "group": "",
+          "help": "Keep molecules with similarity greater or less than the threshold.",
+          "label": "Direction",
+          "options": [
+            "greater",
+            "less"
+          ],
+          "order": 4,
+          "placeholder": "",
+          "widget": "select"
+        },
+        {
+          "default": null,
+          "field": "threshold",
+          "group": "",
+          "help": "Optional similarity bound: results are filtered to this direction.",
+          "label": "Similarity threshold",
+          "options": [],
+          "order": 5,
+          "placeholder": "e.g. 0.8",
+          "widget": "number"
+        },
+        {
+          "default": null,
+          "field": "threads",
+          "group": "",
+          "help": "Threads used to fingerprint and score one shard (default: CPU count).",
+          "label": "Threads per shard",
+          "options": [],
+          "order": 6,
+          "placeholder": "auto",
+          "widget": "number"
+        }
+      ],
+      "upload_ready": true,
+      "verifier": "exact-artifact@1",
+      "version": "1.0.0"
     }
   ]
 }

pyproject.toml

@@ -23,6 +23,7 @@ scimesh = "scimesh.cli:main"
 
 [project.entry-points."scimesh.workloads"]
 "similarity-search@1.0.0" = "scimesh.workloads.search:workload_definition"
+"similarity-search-parallel@1.0.0" = "scimesh.workloads.search_parallel:workload_definition"
 "similarity-graph@1.0.0" = "scimesh.workloads.graph:workload_definition"
 "descriptor-batch@1.0.0" = "scimesh.workloads.descriptors:workload_definition"
 "molwt-filter@1.0.0" = "scimesh.workloads.molwt_filter:workload_definition"

scimesh/workloads/library.py

@@ -21,6 +21,7 @@ from .environment import current_environment_digest
 from .graph import similarity_graph_sdk_definition
 from .molwt_filter import molwt_filter_sdk_definition
 from .search import similarity_search_sdk_definition
+from .search_parallel import similarity_search_parallel_sdk_definition
 
 __all__ = [
     "default_sdk_registry",
@@ -46,6 +47,10 @@ def default_sdk_registry(
         similarity_search_sdk_definition(shard_rows=shard_rows).definition(),
         enabled=True,
     )
+    registry.register(
+        similarity_search_parallel_sdk_definition(shard_rows=shard_rows).definition(),
+        enabled=True,
+    )
     registry.register(
         similarity_graph_sdk_definition().definition(),
         enabled=True,
@@ -82,6 +87,7 @@ def default_sdk_runtime(
             workload_capabilities
             or (
                 "similarity-search",
+                "similarity-search-parallel",
                 "similarity-graph",
                 "descriptor-batch",
                 "molwt-filter",

scimesh/workloads/search_parallel/__init__.py

@@ -0,0 +1,28 @@
+"""SDK-built ``similarity-search-parallel`` workload.
+
+Same contract as ``similarity-search`` with a per-shard thread pool. See
+``core.py`` for the parallel scoring core and ``definition.py`` for the
+manifest-backed handlers.
+"""
+
+from .core import (
+    run_search_shard_parallel,
+    search_similar_parallel,
+    write_search_shards,
+)
+from .definition import (
+    MAP_ENTRY_POINT,
+    SimilaritySearchParallelSDKWorkload,
+    similarity_search_parallel_sdk_definition,
+    workload_definition,
+)
+
+__all__ = [
+    "MAP_ENTRY_POINT",
+    "SimilaritySearchParallelSDKWorkload",
+    "similarity_search_parallel_sdk_definition",
+    "workload_definition",
+    "run_search_shard_parallel",
+    "search_similar_parallel",
+    "write_search_shards",
+]

scimesh/workloads/search_parallel/core.py

@@ -0,0 +1,199 @@
+"""Scientific core for the SDK-built ``similarity-search-parallel`` workload.
+
+The exact same semantics as ``similarity-search`` — identical partial format,
+identical bounded merge, byte-identical output — but the per-molecule
+fingerprinting and Tanimoto scoring of one shard run across a thread pool
+(``threads`` parameter, default = CPU count).
+
+Parallelism is confined to the scoring phase: ``ThreadPoolExecutor.map`` keeps
+the input row order, so the results are merged exactly like the sequential
+reference (same ``_HeapEntry`` logic), which makes the output byte-identical
+for every thread count by construction.
+"""
+
+from __future__ import annotations
+
+import heapq
+import os
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+from typing import Mapping
+
+from rdkit import Chem
+from rdkit.Chem import DataStructs
+
+from scimesh.chemistry.dataset import MoleculeRecord, parse_smiles
+from scimesh.chemistry.fingerprints import fingerprint
+from scimesh.workloads.search.core import (
+    run_search_shard,
+    write_search_partial,
+    write_search_shards,
+)
+from scimesh.workloads.similarity_search import (
+    DatasetStats,
+    SearchResult,
+    SimilarityMatch,
+    _HeapEntry,
+    iter_valid_molecules,
+)
+
+
+def search_similar_parallel(
+    tsv_path: Path,
+    query: MoleculeRecord,
+    top_k: int,
+    *,
+    threads: int = 0,
+    max_rows: int | None = None,
+    threshold: float | None = None,
+    threshold_direction: str = "greater",
+) -> SearchResult:
+    """Exact top-k matches with a bounded heap, scored by a thread pool.
+
+    Identical selection and ordering to ``search_similar`` for every thread
+    count: the merge runs in row order over the parallel-computed scores.
+    """
+    if top_k < 1:
+        raise ValueError("--top-k must be a positive integer")
+    if threads < 0:
+        raise ValueError("threads must be a non-negative integer")
+    if threshold is not None and not 0.0 <= threshold <= 1.0:
+        raise ValueError("--threshold must be between 0 and 1")
+    if threshold_direction not in {"greater", "less"}:
+        raise ValueError("--threshold-direction must be 'greater' or 'less'")
+    workers = threads or (os.cpu_count() or 1)
+
+    query_fingerprint = fingerprint(query.molecule)
+    query_canonical_smiles = Chem.MolToSmiles(query.molecule, canonical=True)
+    stats = DatasetStats()
+    records = list(iter_valid_molecules(tsv_path, stats, max_rows=max_rows))
+
+    def score(record: MoleculeRecord):
+        candidate_smiles = Chem.MolToSmiles(record.molecule, canonical=True)
+        if (
+            record.molecule_id == query.molecule_id
+            or candidate_smiles == query_canonical_smiles
+        ):
+            return None
+        similarity = DataStructs.TanimotoSimilarity(
+            query_fingerprint, fingerprint(record.molecule)
+        )
+        if threshold is not None and (
+            similarity < threshold
+            if threshold_direction == "greater"
+            else similarity > threshold
+        ):
+            return None
+        return similarity
+
+    # map preserves the input order, so the merge below is exactly the
+    # sequential reference's merge, just over precomputed scores.
+    with ThreadPoolExecutor(max_workers=workers) as pool:
+        scored = pool.map(score, records)
+
+    heap: list[_HeapEntry] = []
+    for record, similarity in zip(records, scored):
+        if similarity is None:
+            continue
+        match = SimilarityMatch(similarity, record.molecule_id, record.smiles)
+        rank_key = match.sort_key(threshold_direction)
+        entry = _HeapEntry(match, rank_key)
+        if len(heap) < top_k:
+            heapq.heappush(heap, entry)
+        elif rank_key < heap[0].rank_key:
+            heapq.heapreplace(heap, entry)
+
+    matches = [entry.match for entry in sorted(heap, key=lambda e: e.rank_key)]
+    return SearchResult(matches=matches, stats=stats)
+
+
+def run_search_shard_parallel(
+    input_path: Path,
+    parameters: Mapping[str, object],
+    output_path: Path,
+) -> dict[str, int]:
+    """Run one planned shard with the parallel scoring core.
+
+    Accepts the same parameters as ``similarity-search`` plus ``threads``.
+    """
+    allowed = {
+        "query_id",
+        "query_smiles",
+        "top_k",
+        "threshold",
+        "threshold_direction",
+        "progress_every",
+        "threads",
+    }
+    unknown = set(parameters) - allowed
+    if unknown:
+        raise ValueError(
+            f"unsupported similarity-search-parallel parameters: {', '.join(sorted(unknown))}"
+        )
+    query_smiles = parameters.get("query_smiles")
+    query_id = parameters.get("query_id")
+    if isinstance(query_id, str) and not isinstance(query_smiles, str):
+        from rdkit import Chem
+
+        from scimesh.chemistry.dataset import find_molecule_by_id
+
+        record = find_molecule_by_id(input_path, query_id)
+        query_smiles = Chem.MolToSmiles(record.molecule, canonical=True)
+    if not isinstance(query_smiles, str) or not query_smiles.strip():
+        raise ValueError("query_smiles is required for a distributed shard")
+    molecule = parse_smiles(query_smiles)
+    if molecule is None:
+        raise ValueError("query_smiles is invalid")
+    top_k = _positive_int(parameters.get("top_k", 20), "top_k")
+    threads = _nonnegative_int(parameters.get("threads", 0), "threads")
+    threshold = None
+    if "threshold" in parameters:
+        threshold = _unit_interval(parameters["threshold"], "threshold")
+    direction = parameters.get("threshold_direction", "greater")
+    if direction not in {"greater", "less"}:
+        raise ValueError("threshold_direction must be 'greater' or 'less'")
+    assert isinstance(direction, str)
+
+    result = search_similar_parallel(
+        input_path,
+        MoleculeRecord("query", query_smiles, molecule),
+        top_k=top_k,
+        threads=threads,
+        threshold=threshold,
+        threshold_direction=direction,
+    )
+    write_search_partial(output_path, result.matches)
+    return {
+        "scanned_rows": result.stats.scanned,
+        "valid_molecules": result.stats.valid,
+        "invalid_smiles": result.stats.invalid,
+        "matches_emitted": len(result.matches),
+    }
+
+
+def _positive_int(value: object, name: str) -> int:
+    if isinstance(value, bool) or not isinstance(value, int) or value < 1:
+        raise ValueError(f"{name} must be a positive integer")
+    return value
+
+
+def _nonnegative_int(value: object, name: str) -> int:
+    if isinstance(value, bool) or not isinstance(value, int) or value < 0:
+        raise ValueError(f"{name} must be a non-negative integer")
+    return value
+
+
+def _unit_interval(value: object, name: str) -> float:
+    if isinstance(value, bool) or not isinstance(value, (int, float)):
+        raise ValueError(f"{name} must be a number between 0 and 1")
+    return float(value)
+
+
+# The shared shard writer is re-exported so the workload definition can reuse
+# the deterministic partitioning without importing search internals.
+__all__ = [
+    "search_similar_parallel",
+    "run_search_shard_parallel",
+    "write_search_shards",
+    "run_search_shard",
+]

scimesh/workloads/search_parallel/definition.py

@@ -0,0 +1,144 @@
+"""SDK-built ``similarity-search-parallel`` workload definition and handlers.
+
+A subclass of ``SimilaritySearchSDKWorkload``: identical contract (plan-time
+query resolution, deterministic sharding, top-k reduction, byte-identical
+partials), but each shard's fingerprinting and scoring runs across a thread
+pool (``threads`` parameter, default = CPU count).
+"""
+
+from __future__ import annotations
+
+from collections.abc import Mapping
+from pathlib import Path
+from typing import Any
+
+from scimesh.sdk.batch import MapReduceWorkload
+from scimesh.sdk.identity import WorkloadId
+from scimesh.sdk.plans import JobRequest
+from scimesh.sdk.registry import WorkloadDefinition
+from scimesh.sdk.ui import UIElement
+
+from ..environment import current_environment_digest, current_scimesh_package_digest
+from ..search.definition import SimilaritySearchSDKWorkload, _parameters_schema
+from .core import run_search_shard_parallel
+
+MAP_ENTRY_POINT = "scimesh.workloads.search_parallel.definition:map_search_parallel@v1"
+
+# The parallel variant adds only the thread-count parameter on top of the
+# search contract; everything else (schemas, entry points of the reduce stage,
+# partitioning) is inherited.
+_MAP_PARAMETERS = (
+    "query_id",
+    "query_smiles",
+    "top_k",
+    "threshold",
+    "threshold_direction",
+    "progress_every",
+    "threads",
+)
+
+
+def _parallel_parameters_schema() -> dict[str, Any]:
+    schema = dict(_parameters_schema())
+    properties = dict(schema["properties"])
+    properties["threads"] = {
+        "type": "integer",
+        "minimum": 1,
+        "description": "Threads used to fingerprint and score one shard (default: CPU count).",
+    }
+    schema["properties"] = properties
+    return schema
+
+
+class SimilaritySearchParallelSDKWorkload(SimilaritySearchSDKWorkload):
+    """Exact top-k Tanimoto search with a per-shard thread pool."""
+
+    workload_id = WorkloadId("similarity-search-parallel", "1.0.0")
+    description = (
+        "Exact top-k Tanimoto molecular similarity search over deterministic "
+        "TSV shards with a bounded merge; each shard is fingerprinted and "
+        "scored across a thread pool. Output is byte-identical to "
+        "similarity-search."
+    )
+    parameters_schema = _parallel_parameters_schema()
+    map_parameter_names = _MAP_PARAMETERS
+    map_entry_point = MAP_ENTRY_POINT
+    ui_elements = SimilaritySearchSDKWorkload.ui_elements + (
+        UIElement(
+            "threads",
+            "number",
+            "Threads per shard",
+            help="Threads used to fingerprint and score one shard (default: CPU count).",
+            placeholder="auto",
+            order=6,
+        ),
+    )
+
+    def domain_validate(self, parameters: Mapping[str, Any]) -> None:
+        # The search base rejects unknown parameters; threads is our addition,
+        # so it is validated here and stripped before delegating.
+        rest = dict(parameters)
+        threads = rest.pop("threads", None)
+        if threads is not None and (
+            isinstance(threads, bool) or not isinstance(threads, int) or threads < 1
+        ):
+            raise ValueError("threads must be a positive integer")
+        super().domain_validate(rest)
+
+    def resolved_parameters_for_plan(
+        self,
+        job,
+        input_path,
+        resolved,
+    ):
+        # threads is a map-stage-only knob; strip it from the plan-level
+        # resolved parameters so the reduce stage projection stays clean.
+        resolved = super().resolved_parameters_for_plan(job, input_path, resolved)
+        stripped = dict(resolved)
+        stripped.pop("threads", None)
+        return stripped
+
+    def resolved_parameters(self, request: JobRequest) -> dict[str, Any]:
+        resolved = super().resolved_parameters(request)
+        if "threads" in request.parameters:
+            threads = request.parameters["threads"]
+            if isinstance(threads, bool) or not isinstance(threads, int) or threads < 1:
+                raise ValueError("threads must be a positive integer")
+            resolved["threads"] = threads
+        return resolved
+
+    def compute_shard(
+        self,
+        inputs: Mapping[str, Path],
+        parameters: Mapping[str, Any],
+        output_path: Path,
+    ) -> Mapping[str, int | float]:
+        return run_search_shard_parallel(inputs["input"], parameters, output_path)
+
+
+def similarity_search_parallel_sdk_definition(
+    *,
+    shard_rows: int = 10_000,
+    package_digest: str | None = None,
+    environment_digest: str | None = None,
+) -> SimilaritySearchParallelSDKWorkload:
+    """Build the SDK-built parallel similarity-search definition for tests."""
+    return SimilaritySearchParallelSDKWorkload(
+        shard_rows=shard_rows,
+        package_digest=package_digest or current_scimesh_package_digest(),
+        environment_digest=environment_digest or current_environment_digest(),
+    )
+
+
+def workload_definition() -> WorkloadDefinition:
+    """Installed entry-point factory for the SDK-built parallel search."""
+    return similarity_search_parallel_sdk_definition().definition()
+
+
+def map_search_parallel(
+    input_path: Path,
+    parameters: Mapping[str, object],
+    output_path: Path,
+) -> dict[str, int]:
+    """Digest-pinned map entry point for the parallel search shard."""
+    return run_search_shard_parallel(input_path, parameters, output_path)

tests/test_cli_workload.py

@@ -244,7 +244,13 @@ def test_workload_cli_exports_the_library_as_json(tmp_path: Path) -> None:
     assert payload["schema_version"] == 2
     names = [item["name"] for item in payload["workloads"]]
     assert names == sorted(
-        ["descriptor-batch", "molwt-filter", "similarity-graph", "similarity-search"]
+        [
+            "descriptor-batch",
+            "molwt-filter",
+            "similarity-graph",
+            "similarity-search",
+            "similarity-search-parallel",
+        ]
     )
     for item in payload["workloads"]:
         assert item["version"] == "1.0.0"

tests/test_sdk_compatibility.py

@@ -60,7 +60,7 @@ def _registered_similarity_search(shard_rows: int = 2):
     registry = default_sdk_registry(shard_rows=shard_rows)
     runtime = default_sdk_runtime()
     descriptions = registry.descriptions()
-    assert len(descriptions) == 4
+    assert len(descriptions) == 5
     description = next(
         item for item in descriptions if item.workload.name == "similarity-search"
     )

tests/test_sdk_search_parallel.py

@@ -0,0 +1,185 @@
+"""Tests for the SDK-built similarity-search-parallel workload."""
+
+from __future__ import annotations
+
+import csv
+from pathlib import Path
+
+import pytest
+
+from scimesh.sdk import (
+    ArtifactCollection,
+    DeterminismProfile,
+    JobRequest,
+    LocalArtifactStore,
+    LocalCoreBatchExecutor,
+    LocalPlanningContext,
+    StageKind,
+)
+from scimesh.workloads.library import default_sdk_registry, default_sdk_runtime
+from scimesh.workloads.search.core import run_search_shard, write_search_shards
+from scimesh.workloads.search_parallel import (
+    run_search_shard_parallel,
+    search_similar_parallel,
+)
+from scimesh.workloads.similarity_search import (
+    find_molecule_by_id,
+    search_similar,
+    write_search_results,
+)
+
+
+def _write_dataset(path: Path, molecules: list[tuple[str, str]]) -> None:
+    path.write_text(
+        "chembl_id\tcanonical_smiles\n"
+        + "".join(f"{mid}\t{smiles}\n" for mid, smiles in molecules),
+        encoding="utf-8",
+    )
+
+
+def _tie_dataset(path: Path) -> None:
+    # Deliberate similarity ties: propanol isomers and duplicated rows, so the
+    # parallel merge must reproduce the sequential row-order preference.
+    _write_dataset(
+        path,
+        [
+            ("QUERY", "CCO"),
+            ("A1", "CCCO"),
+            ("A2", "C(CC)O"),
+            ("B", "CCN"),
+            ("C1", "CCC"),
+            ("C2", "CCC"),
+            ("D", "CCCC"),
+        ],
+    )
+
+
+def test_parallel_matches_sequential_byte_exactly(tmp_path: Path) -> None:
+    dataset = tmp_path / "molecules.tsv"
+    _tie_dataset(dataset)
+    query = find_molecule_by_id(dataset, "QUERY")
+
+    reference = search_similar(dataset, query, top_k=5, progress_every=0)
+    reference_path = tmp_path / "reference.csv"
+    write_search_results(reference_path, reference.matches)
+
+    for threads in (1, 2, 4):
+        parallel = search_similar_parallel(dataset, query, top_k=5, threads=threads)
+        parallel_path = tmp_path / f"parallel-{threads}.csv"
+        write_search_results(parallel_path, parallel.matches)
+        assert parallel_path.read_bytes() == reference_path.read_bytes(), (
+            f"threads={threads} diverged from the reference"
+        )
+
+
+def test_parallel_shard_matches_sequential_shard(tmp_path: Path) -> None:
+    dataset = tmp_path / "molecules.tsv"
+    _tie_dataset(dataset)
+    shard_dir = tmp_path / "shards"
+    shard_dir.mkdir()
+    shards = write_search_shards(dataset, shard_dir, shard_rows=2)
+    parameters = {"query_smiles": "CCO", "top_k": 3, "threads": 4}
+
+    sequential_out = tmp_path / "seq.tsv"
+    parallel_out = tmp_path / "par.tsv"
+    sequential_parameters = dict(parameters)
+    sequential_parameters.pop("threads")
+    run_search_shard(shards[0], sequential_parameters, sequential_out)
+    run_search_shard_parallel(shards[0], parameters, parallel_out)
+    assert parallel_out.read_bytes() == sequential_out.read_bytes()
+
+    with parallel_out.open(encoding="utf-8") as handle:
+        rows = list(csv.DictReader(handle))
+    assert rows[0]["rank"] == "1"
+    assert rows[0]["similarity"].startswith("0.5")  # CCO vs CCCO
+    assert len(rows) <= 3
+
+
+def test_parallel_rejects_bad_parameters(tmp_path: Path) -> None:
+    dataset = tmp_path / "molecules.tsv"
+    _tie_dataset(dataset)
+    output = tmp_path / "out.tsv"
+
+    with pytest.raises(ValueError, match="threads must be a non-negative integer"):
+        run_search_shard_parallel(
+            dataset, {"query_smiles": "CCO", "threads": -1}, output
+        )
+    with pytest.raises(ValueError, match="unsupported"):
+        run_search_shard_parallel(dataset, {"query_smiles": "CCO", "nope": 1}, output)
+    with pytest.raises(ValueError, match="query_smiles is invalid"):
+        run_search_shard_parallel(dataset, {"query_smiles": "СС"}, output)
+
+
+def _registered_parallel_search(shard_rows: int = 2):
+    registry = default_sdk_registry(shard_rows=shard_rows)
+    runtime = default_sdk_runtime()
+    description = next(
+        item
+        for item in registry.descriptions()
+        if item.workload.name == "similarity-search-parallel"
+    )
+    definition, negotiated = registry.require(
+        description.workload.name,
+        description.workload.version,
+        description.package_digest,
+        runtime=runtime,
+    )
+    return registry, runtime, description, definition, negotiated
+
+
+def test_parallel_manifest_is_registered_and_negotiable() -> None:
+    _, runtime, description, definition, negotiated = _registered_parallel_search()
+    manifest = definition.manifest
+
+    assert description.enabled is True
+    assert manifest.workload.name == "similarity-search-parallel"
+    assert manifest.workload.version == "1.0.0"
+    assert manifest.determinism is DeterminismProfile.BYTE_EXACT
+    assert manifest.verifier.verifier.canonical == "exact-artifact@1"
+    assert set(mode.value for mode in manifest.trust_modes) == {
+        "trusted",
+        "untrusted_quorum",
+    }
+    assert [stage.kind for stage in manifest.workflow.stages] == [
+        StageKind.MAP,
+        StageKind.REDUCE,
+    ]
+    assert "threads" in manifest.parameters_schema["properties"]
+    assert negotiated is not None
+    assert runtime is not None
+
+
+def test_parallel_executor_matches_reference(tmp_path: Path) -> None:
+    dataset = tmp_path / "molecules.tsv"
+    _tie_dataset(dataset)
+    registry, runtime, description, definition, _ = _registered_parallel_search()
+    artifact_store = LocalArtifactStore(tmp_path / "artifacts")
+    input_port = definition.manifest.inputs["input"]
+    dataset_artifact = artifact_store.import_file(
+        dataset,
+        declaration=input_port.schema,
+    )
+    request = JobRequest(
+        workload=definition.manifest.workload,
+        parameters={"query_id": "QUERY", "top_k": 3, "threads": 2},
+        inputs={"input": ArtifactCollection.single(dataset_artifact)},
+    )
+
+    result = LocalCoreBatchExecutor(
+        registry,
+        runtime,
+        artifact_store,
+        tmp_path / "sdk-work",
+    ).execute(request, description.package_digest)
+    result_artifact = result.outputs["result"].items[0].artifact
+
+    reference_path = tmp_path / "reference.csv"
+    query = find_molecule_by_id(dataset, "QUERY")
+    reference = search_similar(dataset, query, top_k=3, progress_every=0)
+    write_search_results(reference_path, reference.matches)
+
+    assert (
+        artifact_store.materialize(result_artifact).read_bytes()
+        == reference_path.read_bytes()
+    )
+    assert result.task_key == "reduce/final"