Add final result reduction

README.md

@@ -3,10 +3,10 @@
 SciMesh is a scientific-workload framework for molecular datasets. Its public CLI
 runs exact similarity search and sparse similarity-graph construction locally in
 one Python process; it creates no dense similarity matrix. The Go/PostgreSQL
-coordinator and Python worker can run a diagnostic, shard-based
-`similarity-search` pipeline locally. Its CSV artifacts are not a global result
-until CTX-07--09 add planning and reduction; use the local CLI for scientific
-results today. See [`STATUS.md`](STATUS.md).
+coordinator and Python worker can run a shard-based `similarity-search`
+pipeline locally. After every shard succeeds, the coordinator deterministically
+merges its candidates into one final global top-k CSV. See
+[`STATUS.md`](STATUS.md).
 
 The ChEMBL TSV database is intentionally not included in this repository. Download it separately and pass its path to the commands below. The expected columns are `chembl_id` and `canonical_smiles`.
 

STATUS.md

@@ -1,7 +1,7 @@
 # SciMesh Status
 
 **Updated:** 2026-07-24
-**Branch baseline:** `main` at `f953112` (distributed pipeline hardening)
+**Branch baseline:** `main` at `6e67daa` (distributed similarity-search)
 
 ## Current state
 
@@ -18,8 +18,10 @@ the reference behaviour for future distributed execution:
 The Go coordinator and its PostgreSQL-backed task lifecycle are implemented:
 registration, atomic claiming, lease renewal, artifact storage, dataset
 chunking, result/failure reporting, and job progress. The Python worker now
-uses the live coordinator contract; its HTTP path was exercised against a real
-Docker PostgreSQL stack on 2026-07-23.
+uses the live coordinator contract. Completed similarity-search shard results
+are reduced once into a checksum-protected final CSV, which is downloadable
+through the coordinator. The full Go checks (including a fresh migration and
+real PostgreSQL smoke test) passed on 2026-07-24.
 
 ## Milestone tracker
 
@@ -32,29 +34,25 @@ Docker PostgreSQL stack on 2026-07-23.
 | CTX-04 Worker registry and HTTP API | Implemented | Registration, claim, heartbeat, result, failure, and status endpoints. |
 | CTX-05 Artifact storage | Implemented | Coordinator-owned inputs/results, checksum verification, and upload flow. |
 | CTX-06 Python Worker live-contract alignment | Implemented | Worker completed a real uploaded shard via HTTP on 2026-07-23. |
-| CTX-07 Distributed workload protocol | Implemented | Versioned Python contract models, registry, strict plan validation, and deterministic reduction ordering are in `scimesh/distributed/`. The concrete molecular planner/reducer remains CTX-08/09. |
-| CTX-08 Distributed similarity-search | Not started | Local reference exists. |
-| CTX-09 Reducer and final-result API | Not started | Depends on CTX-07 and CTX-08. |
+| CTX-07 Distributed workload protocol | Implemented | Versioned Python contract models, registry, strict plan validation, and deterministic reduction ordering are in `scimesh/distributed/`. |
+| CTX-08 Distributed similarity-search | Implemented | Python planner resolves `query_id` once, creates deterministic shard plans, worker adapter emits exact partial top-k CSVs/metrics, and reducer matches the local reference. |
+| CTX-09 Reducer and final-result API | Implemented | Atomic `reducing` claim, deterministic coordinator-side top-k reducer, sanitized reducer failure, final artifact persistence, `result_uri`, and final CSV download. |
 | CTX-10 Distributed similarity-graph | Not started | Local reference exists. |
-| CTX-11 Dashboard/operator view | Implemented (diagnostic scope) | Protected local view: job/task/worker status, validated similarity-search upload, diagnostic partial-artifact download, and bounded polling. Final-result reduction remains CTX-09. |
+| CTX-11 Dashboard/operator view | Implemented | Protected local view: job/task/worker status, validated similarity-search upload, partial-artifact diagnostics, final-result download, and bounded polling. |
 | CTX-12 Reliability, security, CI | In progress | Unit, race, PostgreSQL integration, and smoke checks exist; CI hardening remains. |
 
 ## Next recommended assignment
 
-Assign **CTX-08** to the workload role: implement the molecular
-`similarity-search` planner and worker adapter on top of the accepted CTX-07
-contract.
+Assign **CTX-10** to the distributed-science role: implement deterministic
+block-pair planning and reduction for `similarity-graph`.
 
 ## Known constraints
 
-- The CTX-07 protocol is implemented, but no concrete molecular planner or
-  reducer is registered yet; the operator UI labels `partial_result` files as
-  diagnostic and cannot present them as final output.
-  Use the local `scimesh` CLI for complete workload results.
 - The worker/coordinator flow currently accepts both underscore API workload
   names and hyphenated CLI names while the contract is consolidated.
-- A real-stack worker test uses a small `query_smiles` shard. Resolving a
-  `query_id` once and sharing it across shards belongs to CTX-07.
+- A real-stack worker test uses a small `query_smiles` shard. The Python
+  planner resolves `query_id` once and shares `query_smiles`; the upload UI
+  currently accepts `query_smiles` only.
 - The coordinator accepts uploaded distributed jobs only for
   `similarity-search` with `query_smiles`. It rejects `similarity-graph` until
   CTX-10 supplies cross-shard pair planning.

coordinator/README.md

@@ -76,9 +76,10 @@ UI_AUTH_TOKEN='local-ui-secret' make up
 ```
 
 The UI is disabled by default and never accepts the worker bearer token.
-It shows recent jobs, task/worker state, and the per-job partial artifacts.
-Those files are explicitly diagnostic until the CTX-09 reducer creates a final
-result; the UI does not present them as final scientific output.
+It shows recent jobs, task/worker state, diagnostic shard artifacts, and the
+final CSV for completed similarity-search jobs. The coordinator enters
+`reducing` after the last shard completes, then exposes the final deterministic
+global top-k result when merging succeeds.
 
 `up` starts three services in order: Postgres waits until `pg_isready` passes, a
 one-shot `migrate` container applies the schema and exits, and only then does the

coordinator/cmd/coordinator/main.go

@@ -75,11 +75,13 @@ func run() error {
 		ClaimTask:        usecase.NewClaimTask(taskRepo, jobRepo, workerRepo, tx, clk, cfg.LeaseDuration),
 		RenewLease:       usecase.NewRenewLease(taskRepo, workerRepo, tx, clk, cfg.LeaseDuration),
 		CompleteTask:     usecase.NewCompleteTask(taskRepo, jobRepo, artifactRepo, tx, clk),
+		ReduceJob:        usecase.NewReduceJob(jobRepo, taskRepo, artifactRepo, blobStore, tx, clk),
 		FailTask:         usecase.NewFailTask(taskRepo, jobRepo, tx, clk),
 		GetJobStatus:     usecase.NewGetJobStatus(jobRepo, taskRepo),
 		CancelJob:        usecase.NewCancelJob(jobRepo, taskRepo, tx, clk),
 		UploadArtifact:   usecase.NewUploadArtifact(taskRepo, artifactRepo, blobStore, tx, clk),
 		DownloadArtifact: usecase.NewDownloadArtifact(artifactRepo, blobStore),
+		GetJobResult:     usecase.NewGetJobResult(jobRepo, usecase.NewDownloadArtifact(artifactRepo, blobStore)),
 		GetTaskInput:     usecase.NewGetTaskInput(taskRepo, artifactRepo, blobStore),
 		Dashboard:        usecase.NewDashboard(uiReadRepo),
 	}

coordinator/internal/domain/job.go

@@ -11,6 +11,7 @@ type JobStatus string
 const (
 	JobPending   JobStatus = "pending"
 	JobRunning   JobStatus = "running"
+	JobReducing  JobStatus = "reducing"
 	JobCompleted JobStatus = "completed"
 	JobFailed    JobStatus = "failed"
 	JobCancelled JobStatus = "cancelled"
@@ -18,14 +19,18 @@ const (
 
 // Job is one user submission that fans out into one or more tasks.
 type Job struct {
-	ID              uuid.UUID
-	Workload        string
-	InputURI        string     // external input URI; empty for uploaded datasets
-	InputArtifactID *uuid.UUID // uploaded input artifact; nil for URI submissions
-	Parameters      map[string]any
-	Status          JobStatus
-	CreatedAt       time.Time
-	CompletedAt     *time.Time
+	ID               uuid.UUID
+	Workload         string
+	InputURI         string     // external input URI; empty for uploaded datasets
+	InputArtifactID  *uuid.UUID // uploaded input artifact; nil for URI submissions
+	ResultArtifactID *uuid.UUID
+	Parameters       map[string]any
+	Status           JobStatus
+	CreatedAt        time.Time
+	CompletedAt      *time.Time
+	ReducerStartedAt *time.Time
+	ErrorCode        *string
+	ErrorMessage     *string
 }
 
 // NewUploadedJob builds a job whose input was uploaded to the coordinator. The
@@ -114,6 +119,8 @@ func (p JobProgress) DeriveStatus() JobStatus {
 	switch {
 	case p.Job.Status == JobCancelled:
 		return JobCancelled
+	case p.Job.Status == JobReducing:
+		return JobReducing
 	case p.Total == 0:
 		return JobPending
 	case p.Done == p.Total:

coordinator/internal/memstore/memstore.go

@@ -216,6 +216,51 @@ func (r *JobRepo) UpdateStatus(ctx context.Context, id uuid.UUID, status domain.
 	return nil
 }
 
+func (r *JobRepo) ClaimReduction(_ context.Context, id uuid.UUID, startedAt time.Time) (bool, error) {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	j, ok := r.jobs[id]
+	if !ok {
+		return false, domain.ErrJobNotFound
+	}
+	if j.Status != domain.JobReducing || j.ReducerStartedAt != nil {
+		return false, nil
+	}
+	j.ReducerStartedAt = &startedAt
+	return true, nil
+}
+
+func (r *JobRepo) CompleteWithResult(_ context.Context, id, resultArtifactID uuid.UUID, completedAt time.Time) error {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	j, ok := r.jobs[id]
+	if !ok {
+		return domain.ErrJobNotFound
+	}
+	j.ResultArtifactID = &resultArtifactID
+	j.Status = domain.JobCompleted
+	j.CompletedAt = &completedAt
+	j.ReducerStartedAt = nil
+	j.ErrorCode = nil
+	j.ErrorMessage = nil
+	return nil
+}
+
+func (r *JobRepo) FailReduction(_ context.Context, id uuid.UUID, code, message string, completedAt time.Time) error {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	j, ok := r.jobs[id]
+	if !ok {
+		return domain.ErrJobNotFound
+	}
+	j.Status = domain.JobFailed
+	j.CompletedAt = &completedAt
+	j.ReducerStartedAt = nil
+	j.ErrorCode = &code
+	j.ErrorMessage = &message
+	return nil
+}
+
 // --- WorkerRepo ----------------------------------------------------------
 
 type WorkerRepo struct {

coordinator/internal/reducer/similarity_search.go

@@ -0,0 +1,195 @@
+// Package reducer contains deterministic, coordinator-side result reductions.
+package reducer
+
+import (
+	"bytes"
+	"encoding/csv"
+	"errors"
+	"fmt"
+	"io"
+	"math"
+	"sort"
+	"strconv"
+)
+
+var searchHeader = []string{"rank", "chembl_id", "canonical_smiles", "similarity"}
+
+type similarityMatch struct {
+	similarity float64
+	id         string
+	smiles     string
+}
+
+// ReduceSimilaritySearch streams worker-local top-k CSVs into the exact global
+// top-k. Each partial is validated before it can affect the final artifact.
+func ReduceSimilaritySearch(partials []io.Reader, parameters map[string]any) ([]byte, error) {
+	topK, err := positiveInt(parameters["top_k"], 20)
+	if err != nil {
+		return nil, err
+	}
+	direction, err := thresholdDirection(parameters["threshold_direction"])
+	if err != nil {
+		return nil, err
+	}
+	h := &matchHeap{direction: direction}
+	for _, partial := range partials {
+		if err := readPartial(partial, direction, func(match similarityMatch) {
+			if len(h.items) < topK {
+				heapPush(h, match)
+				return
+			}
+			if better(match, h.items[0], direction) {
+				h.items[0] = match
+				heapDown(h, 0)
+			}
+		}); err != nil {
+			return nil, err
+		}
+	}
+
+	matches := append([]similarityMatch(nil), h.items...)
+	sort.Slice(matches, func(i, j int) bool { return better(matches[i], matches[j], direction) })
+	var out bytes.Buffer
+	writer := csv.NewWriter(&out)
+	if err := writer.Write(searchHeader); err != nil {
+		return nil, err
+	}
+	for index, match := range matches {
+		if err := writer.Write([]string{
+			strconv.Itoa(index + 1), match.id, match.smiles, fmt.Sprintf("%.6f", match.similarity),
+		}); err != nil {
+			return nil, err
+		}
+	}
+	writer.Flush()
+	if err := writer.Error(); err != nil {
+		return nil, err
+	}
+	return out.Bytes(), nil
+}
+
+func readPartial(input io.Reader, direction string, consume func(similarityMatch)) error {
+	reader := csv.NewReader(input)
+	header, err := reader.Read()
+	if err != nil {
+		return fmt.Errorf("read partial header: %w", err)
+	}
+	if !equalStrings(header, searchHeader) {
+		return fmt.Errorf("partial result has an invalid CSV header")
+	}
+	var previous *similarityMatch
+	for rank := 1; ; rank++ {
+		row, err := reader.Read()
+		if errors.Is(err, io.EOF) {
+			return nil
+		}
+		if err != nil {
+			return fmt.Errorf("read partial row: %w", err)
+		}
+		if len(row) != len(searchHeader) || row[0] != strconv.Itoa(rank) {
+			return fmt.Errorf("partial result has an invalid rank")
+		}
+		score, err := strconv.ParseFloat(row[3], 64)
+		if err != nil || math.IsNaN(score) || math.IsInf(score, 0) || score < 0 || score > 1 {
+			return fmt.Errorf("partial result has an invalid similarity")
+		}
+		match := similarityMatch{similarity: score, id: row[1], smiles: row[2]}
+		if previous != nil && better(match, *previous, direction) {
+			return fmt.Errorf("partial result is not sorted deterministically")
+		}
+		previous = &match
+		consume(match)
+	}
+}
+
+func positiveInt(value any, fallback int) (int, error) {
+	if value == nil {
+		return fallback, nil
+	}
+	switch n := value.(type) {
+	case int:
+		if n > 0 {
+			return n, nil
+		}
+	case int64:
+		if n > 0 && n <= math.MaxInt {
+			return int(n), nil
+		}
+	case float64:
+		if n > 0 && n == math.Trunc(n) && n <= math.MaxInt {
+			return int(n), nil
+		}
+	}
+	return 0, fmt.Errorf("top_k must be a positive integer")
+}
+
+func thresholdDirection(value any) (string, error) {
+	if value == nil {
+		return "greater", nil
+	}
+	direction, ok := value.(string)
+	if !ok || (direction != "greater" && direction != "less") {
+		return "", fmt.Errorf("threshold_direction must be greater or less")
+	}
+	return direction, nil
+}
+
+func better(left, right similarityMatch, direction string) bool {
+	if left.similarity != right.similarity {
+		if direction == "less" {
+			return left.similarity < right.similarity
+		}
+		return left.similarity > right.similarity
+	}
+	if left.id != right.id {
+		return left.id < right.id
+	}
+	return left.smiles < right.smiles
+}
+
+func equalStrings(left, right []string) bool {
+	if len(left) != len(right) {
+		return false
+	}
+	for index := range left {
+		if left[index] != right[index] {
+			return false
+		}
+	}
+	return true
+}
+
+// matchHeap keeps the worst retained match at index zero.
+type matchHeap struct {
+	items     []similarityMatch
+	direction string
+}
+
+func heapPush(h *matchHeap, value similarityMatch) {
+	h.items = append(h.items, value)
+	for child := len(h.items) - 1; child > 0; {
+		parent := (child - 1) / 2
+		if !better(h.items[parent], h.items[child], h.direction) {
+			break
+		}
+		h.items[parent], h.items[child] = h.items[child], h.items[parent]
+		child = parent
+	}
+}
+
+func heapDown(h *matchHeap, parent int) {
+	for {
+		child := parent*2 + 1
+		if child >= len(h.items) {
+			return
+		}
+		if right := child + 1; right < len(h.items) && better(h.items[child], h.items[right], h.direction) {
+			child = right
+		}
+		if !better(h.items[parent], h.items[child], h.direction) {
+			return
+		}
+		h.items[parent], h.items[child] = h.items[child], h.items[parent]
+		parent = child
+	}
+}

coordinator/internal/reducer/similarity_search_test.go

@@ -0,0 +1,41 @@
+package reducer
+
+import (
+	"io"
+	"strings"
+	"testing"
+)
+
+func TestReduceSimilaritySearchKeepsExactCrossShardRanking(t *testing.T) {
+	first := strings.NewReader("rank,chembl_id,canonical_smiles,similarity\n1,B,CCC,0.50000048\n2,C,CCCC,0.1\n")
+	second := strings.NewReader("rank,chembl_id,canonical_smiles,similarity\n1,A,CC,0.50000049\n")
+
+	output, err := ReduceSimilaritySearch([]io.Reader{first, second}, map[string]any{"top_k": 2})
+	if err != nil {
+		t.Fatal(err)
+	}
+	want := "rank,chembl_id,canonical_smiles,similarity\n1,A,CC,0.500000\n2,B,CCC,0.500000\n"
+	if string(output) != want {
+		t.Fatalf("output = %q, want %q", output, want)
+	}
+}
+
+func TestReduceSimilaritySearchSupportsLeastSimilarDirection(t *testing.T) {
+	partial := strings.NewReader("rank,chembl_id,canonical_smiles,similarity\n1,A,CC,0.1\n2,B,CCC,0.8\n")
+	output, err := ReduceSimilaritySearch([]io.Reader{partial}, map[string]any{
+		"top_k": 1, "threshold_direction": "less",
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got, want := string(output), "rank,chembl_id,canonical_smiles,similarity\n1,A,CC,0.100000\n"; got != want {
+		t.Fatalf("output = %q, want %q", got, want)
+	}
+}
+
+func TestReduceSimilaritySearchRejectsMalformedPartial(t *testing.T) {
+	partial := strings.NewReader("rank,chembl_id,canonical_smiles,similarity\n2,A,CC,0.1\n")
+	if _, err := ReduceSimilaritySearch([]io.Reader{partial}, nil); err == nil {
+		t.Fatal("expected malformed rank error")
+	}
+}

coordinator/internal/storage/postgres/integration_test.go

@@ -90,6 +90,49 @@ func TestCreateJobPersistsEveryTask(t *testing.T) {
 	}
 }
 
+func TestClaimReductionIsAtomic(t *testing.T) {
+	pool := testPool(t)
+	job, _ := seedJob(t, pool, 1)
+	repo := NewJobRepo(pool)
+	ctx := context.Background()
+	if err := repo.UpdateStatus(ctx, job.ID, domain.JobReducing, nil); err != nil {
+		t.Fatal(err)
+	}
+
+	var (
+		wg      sync.WaitGroup
+		mu      sync.Mutex
+		claimed int
+	)
+	for range 8 {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			ok, err := repo.ClaimReduction(context.Background(), job.ID, time.Now().UTC())
+			if err != nil {
+				t.Errorf("claim reduction: %v", err)
+				return
+			}
+			if ok {
+				mu.Lock()
+				claimed++
+				mu.Unlock()
+			}
+		}()
+	}
+	wg.Wait()
+	if claimed != 1 {
+		t.Fatalf("reducer claims = %d, want 1", claimed)
+	}
+	stored, err := repo.Get(ctx, job.ID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if stored.Status != domain.JobReducing || stored.ReducerStartedAt == nil {
+		t.Fatalf("stored reduction state = %+v", stored)
+	}
+}
+
 // A job must land whole or not at all: a half-created job leaves chunks no
 // worker could ever complete.
 func TestCreateJobRollsBackOnFailure(t *testing.T) {

coordinator/internal/storage/postgres/job_repo.go

@@ -25,7 +25,10 @@ func NewJobRepo(pool *pgxpool.Pool) *JobRepo {
 
 var _ usecase.JobRepository = (*JobRepo)(nil)
 
-var jobColumns = []string{"id", "workload", "input_uri", "parameters", "status", "created_at", "completed_at"}
+var jobColumns = []string{
+	"id", "workload", "input_uri", "parameters", "status", "created_at", "completed_at",
+	"input_artifact_id", "result_artifact_id", "error_code", "error_message", "reducer_started_at",
+}
 
 // Insert runs inside the caller's transaction, alongside the job's tasks — that
 // is what makes "all tasks or none" hold.
@@ -55,7 +58,8 @@ func (r *JobRepo) Get(ctx context.Context, id uuid.UUID) (*domain.Job, error) {
 		status string
 	)
 	err = conn(ctx, r.pool).QueryRow(ctx, sql, args...).Scan(
-		&j.ID, &j.Workload, &j.InputURI, &j.Parameters, &status, &j.CreatedAt, &j.CompletedAt)
+		&j.ID, &j.Workload, &j.InputURI, &j.Parameters, &status, &j.CreatedAt, &j.CompletedAt,
+		&j.InputArtifactID, &j.ResultArtifactID, &j.ErrorCode, &j.ErrorMessage, &j.ReducerStartedAt)
 	if errors.Is(err, pgx.ErrNoRows) {
 		return nil, domain.ErrJobNotFound
 	}
@@ -66,6 +70,64 @@ func (r *JobRepo) Get(ctx context.Context, id uuid.UUID) (*domain.Job, error) {
 	return &j, nil
 }
 
+func (r *JobRepo) ClaimReduction(ctx context.Context, id uuid.UUID, startedAt time.Time) (bool, error) {
+	sql, args, err := psql.Update("jobs").
+		Set("reducer_started_at", startedAt).
+		Where(sq.Eq{"id": id, "status": string(domain.JobReducing), "reducer_started_at": nil}).
+		ToSql()
+	if err != nil {
+		return false, err
+	}
+	tag, err := conn(ctx, r.pool).Exec(ctx, sql, args...)
+	if err != nil {
+		return false, err
+	}
+	return tag.RowsAffected() == 1, nil
+}
+
+func (r *JobRepo) CompleteWithResult(ctx context.Context, id, resultArtifactID uuid.UUID, completedAt time.Time) error {
+	sql, args, err := psql.Update("jobs").
+		SetMap(map[string]any{
+			"status":             string(domain.JobCompleted),
+			"result_artifact_id": resultArtifactID,
+			"completed_at":       completedAt,
+			"reducer_started_at": nil,
+			"error_code":         nil,
+			"error_message":      nil,
+		}).
+		Where(sq.Eq{"id": id, "status": string(domain.JobReducing)}).
+		ToSql()
+	if err != nil {
+		return err
+	}
+	tag, err := conn(ctx, r.pool).Exec(ctx, sql, args...)
+	if err != nil {
+		return err
+	}
+	if tag.RowsAffected() == 0 {
+		return domain.ErrJobNotFound
+	}
+	return nil
+}
+
+func (r *JobRepo) FailReduction(ctx context.Context, id uuid.UUID, code, message string, completedAt time.Time) error {
+	sql, args, err := psql.Update("jobs").
+		SetMap(map[string]any{
+			"status":             string(domain.JobFailed),
+			"completed_at":       completedAt,
+			"error_code":         code,
+			"error_message":      message,
+			"reducer_started_at": nil,
+		}).
+		Where(sq.Eq{"id": id, "status": string(domain.JobReducing)}).
+		ToSql()
+	if err != nil {
+		return err
+	}
+	_, err = conn(ctx, r.pool).Exec(ctx, sql, args...)
+	return err
+}
+
 func (r *JobRepo) UpdateStatus(ctx context.Context, id uuid.UUID,
 	status domain.JobStatus, completedAt *time.Time) error {
 

coordinator/internal/transport/http/dto.go

@@ -119,6 +119,8 @@ type jobProgressResponse struct {
 	Done      int       `json:"completed"`
 	Failed    int       `json:"failed"`
 	Cancelled int       `json:"cancelled"`
+	ResultURI string    `json:"result_uri,omitempty"`
+	ErrorCode string    `json:"error_code,omitempty"`
 }
 
 type uploadArtifactResponse struct {
@@ -153,7 +155,7 @@ func toClaimedTaskResponse(c domain.ClaimedTask) claimedTaskResponse {
 }
 
 func toJobProgressResponse(p domain.JobProgress) jobProgressResponse {
-	return jobProgressResponse{
+	out := jobProgressResponse{
 		ID:        p.Job.ID,
 		Status:    string(p.DeriveStatus()),
 		Total:     p.Total,
@@ -163,4 +165,11 @@ func toJobProgressResponse(p domain.JobProgress) jobProgressResponse {
 		Failed:    p.Failed,
 		Cancelled: p.Cancelled,
 	}
+	if p.Job.ResultArtifactID != nil && out.Status == string(domain.JobCompleted) {
+		out.ResultURI = "/jobs/" + p.Job.ID.String() + "/result"
+	}
+	if p.Job.ErrorCode != nil {
+		out.ErrorCode = *p.Job.ErrorCode
+	}
+	return out
 }

coordinator/internal/transport/http/dto_test.go

@@ -0,0 +1,25 @@
+package http
+
+import (
+	"testing"
+
+	"github.com/google/uuid"
+
+	"github.com/emil28092005/SciMesh/coordinator/internal/domain"
+)
+
+func TestJobProgressResponseExposesFinalResultOnlyWhenCompleted(t *testing.T) {
+	id := uuid.New()
+	result := uuid.New()
+	progress := domain.JobProgress{Job: domain.Job{
+		ID: id, Status: domain.JobCompleted, ResultArtifactID: &result,
+	}, Total: 1, Done: 1}
+	if got, want := toJobProgressResponse(progress).ResultURI, "/jobs/"+id.String()+"/result"; got != want {
+		t.Fatalf("result URI = %q, want %q", got, want)
+	}
+
+	progress.Job.Status = domain.JobReducing
+	if got := toJobProgressResponse(progress).ResultURI; got != "" {
+		t.Fatalf("reducing job exposes result URI %q", got)
+	}
+}

coordinator/internal/transport/http/handlers.go

@@ -150,6 +150,12 @@ func (s *Server) handleResult(w http.ResponseWriter, r *http.Request) {
 		s.writeError(w, r, err)
 		return
 	}
+	if s.uc.ReduceJob != nil {
+		if err := s.uc.ReduceJob.Execute(ctx, task.JobID); err != nil {
+			s.writeError(w, r, err)
+			return
+		}
+	}
 	writeJSON(w, http.StatusOK, taskResponse{ID: task.ID, JobID: task.JobID, Status: string(task.Status)})
 }
 
@@ -399,6 +405,25 @@ func (s *Server) handleGetJob(w http.ResponseWriter, r *http.Request) {
 	writeJSON(w, http.StatusOK, toJobProgressResponse(progress))
 }
 
+func (s *Server) handleGetJobResult(w http.ResponseWriter, r *http.Request) {
+	ctx, cancel := s.reqCtx(r)
+	defer cancel()
+	jobID, ok := s.pathUUID(w, r, "job_id")
+	if !ok {
+		return
+	}
+	art, body, err := s.uc.GetJobResult.Execute(ctx, jobID)
+	if err != nil {
+		s.writeError(w, r, err)
+		return
+	}
+	defer func() { _ = body.Close() }()
+	w.Header().Set("Content-Type", art.ContentType)
+	w.Header().Set("Content-Length", strconv.FormatInt(art.SizeBytes, 10))
+	w.Header().Set("Content-Disposition", fmt.Sprintf("attachment; filename=%q", art.Filename))
+	_, _ = io.Copy(w, body)
+}
+
 // handleCancelJob stops all non-terminal shards for an operator-requested job.
 // It is available to both the bearer API and the separately authenticated UI.
 func (s *Server) handleCancelJob(w http.ResponseWriter, r *http.Request) {

coordinator/internal/transport/http/server.go

@@ -22,8 +22,10 @@ type UseCases struct {
 	ClaimTask        *usecase.ClaimTask
 	RenewLease       *usecase.RenewLease
 	CompleteTask     *usecase.CompleteTask
+	ReduceJob        *usecase.ReduceJob
 	FailTask         *usecase.FailTask
 	GetJobStatus     *usecase.GetJobStatus
+	GetJobResult     *usecase.GetJobResult
 	CancelJob        *usecase.CancelJob
 	UploadArtifact   *usecase.UploadArtifact
 	DownloadArtifact *usecase.DownloadArtifact
@@ -62,6 +64,7 @@ func (s *Server) Handler(token string, uiToken ...string) http.Handler {
 	protected.HandleFunc("POST /jobs", s.handleCreateJob)
 	protected.HandleFunc("POST /jobs/upload", s.handleUploadDataset)
 	protected.HandleFunc("GET /jobs/{job_id}", s.handleGetJob)
+	protected.HandleFunc("GET /jobs/{job_id}/result", s.handleGetJobResult)
 	protected.HandleFunc("POST /jobs/{job_id}/cancel", s.handleCancelJob)
 	protected.HandleFunc("POST /tasks/claim", s.handleClaim)
 	protected.HandleFunc("GET /tasks/{task_id}/input", s.handleGetTaskInput)

coordinator/internal/transport/http/server_test.go

@@ -42,6 +42,7 @@ func newEnvWithUIToken(t *testing.T, ready func(context.Context) error, configur
 	clk := memstore.NewClock(time.Date(2026, 7, 21, 12, 0, 0, 0, time.UTC))
 	tx := memstore.Tx{}
 	lease := 2 * time.Minute
+	downloadArtifact := usecase.NewDownloadArtifact(arts, blobs)
 
 	uc := coordhttp.UseCases{
 		RegisterWorker:   usecase.NewRegisterWorker(work, clk),
@@ -50,11 +51,13 @@ func newEnvWithUIToken(t *testing.T, ready func(context.Context) error, configur
 		ClaimTask:        usecase.NewClaimTask(tasks, jobs, work, tx, clk, lease),
 		RenewLease:       usecase.NewRenewLease(tasks, work, tx, clk, lease),
 		CompleteTask:     usecase.NewCompleteTask(tasks, jobs, arts, tx, clk),
+		ReduceJob:        usecase.NewReduceJob(jobs, tasks, arts, blobs, tx, clk),
 		FailTask:         usecase.NewFailTask(tasks, jobs, tx, clk),
 		GetJobStatus:     usecase.NewGetJobStatus(jobs, tasks),
+		GetJobResult:     usecase.NewGetJobResult(jobs, downloadArtifact),
 		CancelJob:        usecase.NewCancelJob(jobs, tasks, tx, clk),
 		UploadArtifact:   usecase.NewUploadArtifact(tasks, arts, blobs, tx, clk),
-		DownloadArtifact: usecase.NewDownloadArtifact(arts, blobs),
+		DownloadArtifact: downloadArtifact,
 		GetTaskInput:     usecase.NewGetTaskInput(tasks, arts, blobs),
 		Dashboard:        usecase.NewDashboard(memstore.NewUIReadRepo(jobs, tasks, work, arts)),
 	}
@@ -400,6 +403,42 @@ func TestFullLifecycle(t *testing.T) {
 	}
 }
 
+func TestSimilaritySearchLifecyclePublishesFinalResult(t *testing.T) {
+	e := newEnv(t, healthy)
+	code, job := e.uploadDataset(t, "similarity-search", 10, "chembl_id\tcanonical_smiles\nA\tCC\n")
+	if code != http.StatusCreated {
+		t.Fatalf("upload job: %d", code)
+	}
+	jobID := job["job_id"].(string)
+	code, claim := e.do(t, "POST", "/tasks/claim", `{"worker_id":"w1","capabilities":["similarity-search"]}`)
+	if code != http.StatusOK {
+		t.Fatalf("claim: %d", code)
+	}
+	taskID := claim["task_id"].(string)
+	attempt := int(claim["attempt"].(float64))
+	artifactID := e.putArtifact(t, taskID, "w1", attempt, "rank,chembl_id,canonical_smiles,similarity\n1,A,CC,0.900000\n")
+	if code, _ := e.do(t, "POST", "/tasks/"+taskID+"/result",
+		`{"worker_id":"w1","attempt":`+itoa(attempt)+`,"result":{"artifact_id":"`+artifactID+`"}}`); code != http.StatusOK {
+		t.Fatalf("complete: %d", code)
+	}
+
+	code, progress := e.do(t, "GET", "/jobs/"+jobID, "")
+	if code != http.StatusOK || progress["status"] != "completed" || progress["result_uri"] != "/jobs/"+jobID+"/result" {
+		t.Fatalf("progress = (%d, %v)", code, progress)
+	}
+	req, _ := http.NewRequestWithContext(context.Background(), "GET", e.ts.URL+progress["result_uri"].(string), nil)
+	req.Header.Set("Authorization", "Bearer "+token)
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	body, _ := io.ReadAll(resp.Body)
+	if resp.StatusCode != http.StatusOK || string(body) != "rank,chembl_id,canonical_smiles,similarity\n1,A,CC,0.900000\n" {
+		t.Fatalf("final result = (%d, %q)", resp.StatusCode, body)
+	}
+}
+
 func TestForeignArtifactResultConflict(t *testing.T) {
 	e := newEnv(t, healthy)
 	e.do(t, "POST", "/jobs", `{"workload":"w","input_uri":"s3://in",

coordinator/internal/transport/http/templates/dashboard.html

@@ -12,7 +12,7 @@
 <body>
   <main class="page">
     <header class="top"><div><p class="eyebrow">Local coordinator</p><h1>SciMesh operator dashboard</h1><p class="lead">See where a computation is and what should happen next.</p></div><a class="button" href="/ui/jobs/new">Start a check</a></header>
-    <section class="notice" aria-label="Current pipeline limitation"><strong>This screen currently diagnoses shard jobs.</strong><span>Workers upload partial CSVs to the coordinator. Until a reducer is implemented, those files are not one final scientific result.</span><div class="steps"><div class="step"><b>1. Upload TSV</b>The coordinator splits the file into shard tasks.</div><div class="step"><b>2. Wait for a worker</b>A worker claims a shard, calculates similarity, and returns a CSV.</div><div class="step"><b>3. Inspect artifacts</b>Download a partial result from the job page.</div></div></section>
+    <section class="notice" aria-label="Pipeline flow"><strong>Similarity-search jobs produce a final CSV.</strong><span>Workers return shard-level candidates; after every shard succeeds, the coordinator deterministically merges them into one global top-k result.</span><div class="steps"><div class="step"><b>1. Upload TSV</b>The coordinator splits the file into shard tasks.</div><div class="step"><b>2. Wait for a worker</b>A worker claims a shard, calculates similarity, and returns a CSV.</div><div class="step"><b>3. Download result</b>When merging finishes, download the final CSV from the job page.</div></div></section>
     <h2>Recent jobs</h2>
     <div class="table-wrap"><table><tr><th>Computation</th><th>State</th><th>Progress</th><th class="hide-mobile">Created</th></tr>{{range .Jobs}}<tr><td><a href="/ui/jobs/{{.ID}}"><strong>{{workloadLabel .Workload}}</strong></a><br><small>Open job details</small></td><td><span class="status status-{{statusClass .Status}}">{{statusLabel .Status}}</span><br><small>{{statusHint .Status}}</small></td><td class="kicker"><strong>{{.Completed}} / {{.Total}}</strong> complete{{if gt .Failed 0}} · <span style="color:#a31135">failed: {{.Failed}}</span>{{end}}{{if gt .Cancelled 0}} · <span>stopped: {{.Cancelled}}</span>{{end}}<div class="bar"><span style="width:{{progressPercent .Completed .Failed .Cancelled .Total}}%"></span></div></td><td class="hide-mobile"><small>{{time .CreatedAt}}</small></td></tr>{{else}}<tr><td colspan="4" class="empty"><strong>No jobs yet.</strong><br>Click “Start a check”, upload a small TSV, and leave a worker running.</td></tr>{{end}}</table></div>
     <h2>Workers</h2>

coordinator/internal/transport/http/templates/job.html

@@ -13,14 +13,14 @@
   <main class="page">
     <a class="back" href="/ui">← Back to jobs</a><p class="eyebrow">{{workloadLabel .Workload}}</p><h1>Execution progress</h1>
     <section class="summary"><div class="summary-head"><div><span id="status" class="status status-{{statusClass .Status}}">{{statusLabel .Status}}</span><p id="hint" class="hint">{{statusHint .Status}}</p></div><div>{{if cancellable .Status}}<button id="stop-job" class="stop" type="button">Stop job</button><small>This cancels every shard that is not finished yet.</small>{{else}}<small>Summary refreshes automatically every two seconds.</small>{{end}}</div></div><div class="bar" aria-label="Progress"><span id="progress-bar" style="width:{{progressPercent .Completed .Failed .Cancelled .Total}}%"></span></div><p id="progress" class="muted">{{.Completed}} of {{.Total}} tasks complete</p><div class="numbers"><div class="number"><b id="total">{{.Total}}</b><small>total shards</small></div><div class="number"><b id="completed">{{.Completed}}</b><small>complete</small></div><div class="number"><b id="pending">{{.Pending}}</b><small>waiting</small></div><div class="number"><b id="active">{{add .Leased .Running}}</b><small>with workers</small></div><div class="number"><b id="failed">{{.Failed}}</b><small>failed</small></div><div class="number"><b id="cancelled">{{.Cancelled}}</b><small>stopped</small></div></div><details><summary>Technical details</summary><p>Job ID: <code>{{.ID}}</code><br>Workload: <code>{{.Workload}}</code><br>Created: {{time .CreatedAt}}</p></details></section>
-    <section class="notice"><strong>What can be downloaded now?</strong><br><code>partial_result</code> files come from individual shards. They are useful for checking the pipeline, but are not a merged final CSV because the reducer is not implemented yet.</section>
+    <section class="notice">{{if .FinalResultAvailable}}<strong>Final result ready</strong><br>Download the <code>final_result</code> CSV below. It is the deterministic global top-k across all completed shards. The <code>partial_result</code> files remain available for diagnostics.{{else if eq .Status "reducing"}}<strong>Merging completed shards</strong><br>Every shard has finished. The coordinator is building one deterministic global CSV; refresh in a moment to download it.{{else}}<strong>What can be downloaded now?</strong><br><code>partial_result</code> files come from individual shards. They are useful for checking the pipeline, but are not a merged final CSV yet.{{end}}</section>
     <h2>Shard tasks</h2><p class="muted">If a task fails, its code and message appear here. Refresh the page to update the detailed rows.</p>
     <div class="table-wrap"><table><tr><th>Shard</th><th>State</th><th>Attempt</th><th>Worker / lease</th><th>Error</th></tr>{{range .Tasks}}<tr><td>#{{.ChunkIndex}}</td><td><span class="status status-{{statusClass .Status}}">{{statusLabel .Status}}</span></td><td>{{.Attempt}} / {{.MaxAttempts}}</td><td>{{if .LeaseOwner}}<code>{{.LeaseOwner}}</code>{{if .LeaseExpiresAt}}<br><small>until {{time .LeaseExpiresAt}}</small>{{end}}{{else}}<span class="muted">—</span>{{end}}</td><td class="error">{{if .ErrorCode}}<strong>{{taskErrorLabel .ErrorCode}}</strong><br><small>{{taskErrorHint .ErrorCode}}</small>{{else}}<span class="muted">—</span>{{end}}</td></tr>{{else}}<tr><td colspan="5" class="empty">No tasks have appeared yet.</td></tr>{{end}}</table></div>
     <h2>Coordinator artifacts</h2>
     <div class="table-wrap"><table><tr><th>Type</th><th>File</th><th>Size</th><th>Integrity check</th><th></th></tr>{{range .Artifacts}}<tr><td>{{if .Diagnostic}}<strong>Partial result</strong><br><small>diagnostic</small>{{else}}{{.Kind}}{{end}}</td><td>{{.Filename}}</td><td>{{bytes .SizeBytes}}</td><td><code>{{.SHA256}}</code></td><td>{{if .Downloadable}}<a class="download" href="/ui/jobs/{{$.ID}}/artifacts/{{.ID}}">Download CSV</a>{{else}}<span class="muted">Unavailable</span>{{end}}</td></tr>{{else}}<tr><td colspan="5" class="empty">No artifacts yet. The worker uploads a CSV after it completes a shard.</td></tr>{{end}}</table></div>
   </main>
   <script>
-    const id={{printf "%q" .ID}},state={pending:['Waiting for a worker','waiting','Waiting for an available worker with the required capability.'],leased:['Assigned to a worker','active','A worker has claimed the task and should begin processing shortly.'],running:['Running','active','A worker is reading a shard, calculating fingerprints, and uploading its result through the coordinator.'],completed:['Tasks complete','success','Every shard task is complete. Files below are still partial results.'],failed:['Needs attention','danger','One or more shard tasks failed. Open the task list below for details.'],cancelled:['Stopped','waiting','The operator stopped this job. No new shards can be claimed.']};
+    const id={{printf "%q" .ID}},state={pending:['Waiting for a worker','waiting','Waiting for an available worker with the required capability.'],leased:['Assigned to a worker','active','A worker has claimed the task and should begin processing shortly.'],running:['Running','active','A worker is reading a shard, calculating fingerprints, and uploading its result through the coordinator.'],reducing:['Merging results','active','All shards are complete. The coordinator is merging their candidates into one final CSV.'],completed:['Completed','success','The final result is ready to download.'],failed:['Needs attention','danger','One or more shard tasks failed. Open the task list below for details.'],cancelled:['Stopped','waiting','The operator stopped this job. No new shards can be claimed.']};
     const stop=document.querySelector('#stop-job');if(stop)stop.addEventListener('click',async()=>{if(!confirm('Stop this job? Unfinished shards will be cancelled.'))return;stop.disabled=true;const response=await fetch('/ui/api/jobs/'+id+'/cancel',{method:'POST'});if(!response.ok){stop.disabled=false;alert('Unable to stop this job.');return}location.reload()});
     const terminal=new Set(['completed','failed','cancelled']);let timer;const poll=async()=>{try{const response=await fetch('/ui/api/jobs/'+id);if(!response.ok)return;const job=await response.json(),info=state[job.status]||[job.status,'waiting','Status reported by the coordinator.'],done=job.completed+job.failed+job.cancelled,percent=job.total?Math.min(100,Math.floor(done*100/job.total)):0,badge=document.querySelector('#status');badge.textContent=info[0];badge.className='status status-'+info[1];document.querySelector('#hint').textContent=info[2];document.querySelector('#progress').textContent=job.completed+' of '+job.total+' tasks complete'+(job.failed?' · failed: '+job.failed:'')+(job.cancelled?' · stopped: '+job.cancelled:'');document.querySelector('#progress-bar').style.width=percent+'%';for(const key of ['total','completed','pending','failed','cancelled'])document.querySelector('#'+key).textContent=job[key];document.querySelector('#active').textContent=job.leased+job.running;if(terminal.has(job.status)&&timer){clearInterval(timer);timer=undefined}}catch(_){}};const start=()=>{if(!timer&&!document.hidden&&!terminal.has(document.querySelector('#status').textContent.toLowerCase()))timer=setInterval(poll,2000)};document.addEventListener('visibilitychange',()=>{if(document.hidden&&timer){clearInterval(timer);timer=undefined}else start()});start();
   </script>

coordinator/internal/transport/http/ui.go

@@ -48,8 +48,10 @@ func uiStatusLabel(status string) string {
 		return "Assigned to a worker"
 	case "running":
 		return "Running"
+	case "reducing":
+		return "Merging results"
 	case "completed":
-		return "Tasks complete"
+		return "Completed"
 	case "failed":
 		return "Needs attention"
 	case "cancelled":
@@ -67,8 +69,10 @@ func uiStatusHint(status string) string {
 		return "A worker has claimed the task and should begin processing shortly."
 	case "running":
 		return "A worker is reading a shard, calculating fingerprints, and uploading its result through the coordinator."
+	case "reducing":
+		return "All shards are complete. The coordinator is merging their candidates into one final CSV."
 	case "completed":
-		return "Every shard task is complete. Files below are still partial results."
+		return "The final result is ready to download."
 	case "failed":
 		return "One or more shard tasks failed. Open the task list below for details."
 	case "cancelled":
@@ -86,7 +90,7 @@ func uiStatusClass(status string) string {
 		return "danger"
 	case "cancelled":
 		return "waiting"
-	case "running", "leased":
+	case "running", "leased", "reducing":
 		return "active"
 	default:
 		return "waiting"

coordinator/internal/transport/http/ui_test.go

@@ -10,7 +10,8 @@ func TestUIStatusPresentation(t *testing.T) {
 	}{
 		{"pending", "Waiting for a worker", "waiting"},
 		{"running", "Running", "active"},
-		{"completed", "Tasks complete", "success"},
+		{"reducing", "Merging results", "active"},
+		{"completed", "Completed", "success"},
 		{"failed", "Needs attention", "danger"},
 	}
 	for _, test := range tests {

coordinator/internal/usecase/job.go

@@ -100,7 +100,7 @@ func (uc *CancelJob) Execute(ctx context.Context, jobID uuid.UUID) (int64, error
 		if job.Status == domain.JobCancelled {
 			return nil
 		}
-		if job.Status == domain.JobCompleted || job.Status == domain.JobFailed {
+		if job.Status == domain.JobReducing || job.Status == domain.JobCompleted || job.Status == domain.JobFailed {
 			return domain.ErrJobNotCancellable
 		}
 		// The lease reaper can be the transition that exhausted the final task.
@@ -111,7 +111,7 @@ func (uc *CancelJob) Execute(ctx context.Context, jobID uuid.UUID) (int64, error
 			return err
 		}
 		derived := progressFrom(*job, counts).DeriveStatus()
-		if derived == domain.JobCompleted || derived == domain.JobFailed {
+		if derived == domain.JobReducing || derived == domain.JobCompleted || derived == domain.JobFailed {
 			return domain.ErrJobNotCancellable
 		}
 		cancelled, err = uc.tasks.CancelByJob(ctx, jobID, now)
@@ -226,10 +226,20 @@ func syncJobStatus(ctx context.Context, jobs JobRepository, tasks TaskRepository
 	if err != nil {
 		return err
 	}
-	status := progressFrom(domain.Job{}, counts).DeriveStatus()
+	job, err := jobs.Get(ctx, jobID)
+	if err != nil {
+		return err
+	}
+	status := progressFrom(*job, counts).DeriveStatus()
+	// All worker shards being complete means scientific reduction is ready, not
+	// that the job's final artifact already exists. CTX-09 owns the transition
+	// from reducing to completed after it persists that artifact.
+	if status == domain.JobCompleted && job.Workload == "similarity-search" {
+		status = domain.JobReducing
+	}
 
 	var completedAt *time.Time
-	if status == domain.JobCompleted || status == domain.JobFailed {
+	if status == domain.JobFailed {
 		completedAt = &now
 	}
 	return jobs.UpdateStatus(ctx, jobID, status, completedAt)

coordinator/internal/usecase/ports.go

@@ -70,6 +70,9 @@ type JobRepository interface {
 	Insert(ctx context.Context, j *domain.Job) error
 	Get(ctx context.Context, id uuid.UUID) (*domain.Job, error)
 	UpdateStatus(ctx context.Context, id uuid.UUID, status domain.JobStatus, completedAt *time.Time) error
+	ClaimReduction(ctx context.Context, id uuid.UUID, startedAt time.Time) (bool, error)
+	CompleteWithResult(ctx context.Context, id, resultArtifactID uuid.UUID, completedAt time.Time) error
+	FailReduction(ctx context.Context, id uuid.UUID, code, message string, completedAt time.Time) error
 }
 
 // WorkerRepository persists the worker registry.

coordinator/internal/usecase/reduce.go

@@ -0,0 +1,133 @@
+package usecase
+
+import (
+	"bytes"
+	"context"
+	"io"
+
+	"github.com/google/uuid"
+
+	"github.com/emil28092005/SciMesh/coordinator/internal/domain"
+	"github.com/emil28092005/SciMesh/coordinator/internal/reducer"
+)
+
+// ReduceJob turns completed coordinator-owned partial artifacts into one final
+// artifact. It performs no worker I/O and never trusts a worker URI or path.
+type ReduceJob struct {
+	jobs      JobRepository
+	tasks     TaskRepository
+	artifacts ArtifactRepository
+	blobs     BlobStore
+	tx        TxManager
+	clock     Clock
+}
+
+func NewReduceJob(jobs JobRepository, tasks TaskRepository, artifacts ArtifactRepository,
+	blobs BlobStore, tx TxManager, clock Clock) *ReduceJob {
+	return &ReduceJob{jobs: jobs, tasks: tasks, artifacts: artifacts, blobs: blobs, tx: tx, clock: clock}
+}
+
+// Execute is idempotent for jobs that are not currently reducing. The worker
+// completion path may call it after every result; only the last task changes a
+// similarity-search job into reducing state.
+func (uc *ReduceJob) Execute(ctx context.Context, jobID uuid.UUID) error {
+	claimed, err := uc.jobs.ClaimReduction(ctx, jobID, uc.clock.Now())
+	if err != nil || !claimed {
+		return err
+	}
+	job, err := uc.jobs.Get(ctx, jobID)
+	if err != nil {
+		return err
+	}
+	if job.Status != domain.JobReducing {
+		return nil
+	}
+	if job.Workload != "similarity-search" {
+		return uc.fail(ctx, jobID)
+	}
+
+	completed, err := uc.tasks.ListCompleted(ctx, jobID)
+	if err != nil {
+		return uc.fail(ctx, jobID)
+	}
+	if len(completed) == 0 {
+		return uc.fail(ctx, jobID)
+	}
+	readers := make([]io.Reader, 0, len(completed))
+	closers := make([]io.Closer, 0, len(completed))
+	for _, task := range completed {
+		if task.ResultArtifactID == nil {
+			closeAll(closers)
+			return uc.fail(ctx, jobID)
+		}
+		artifact, err := uc.artifacts.Get(ctx, *task.ResultArtifactID)
+		if err != nil || artifact.JobID != jobID || artifact.TaskID == nil || *artifact.TaskID != task.ID || artifact.Kind != domain.ArtifactPartialResult {
+			closeAll(closers)
+			return uc.fail(ctx, jobID)
+		}
+		body, err := uc.blobs.Open(ctx, artifact.StorageKey)
+		if err != nil {
+			closeAll(closers)
+			return uc.fail(ctx, jobID)
+		}
+		readers = append(readers, body)
+		closers = append(closers, body)
+	}
+	output, reduceErr := reducer.ReduceSimilaritySearch(readers, job.Parameters)
+	closeAll(closers)
+	if reduceErr != nil {
+		return uc.fail(ctx, jobID)
+	}
+
+	final, err := domain.NewArtifact(jobID, nil, domain.ArtifactFinalResult, "similarity-search.csv", "text/csv", uc.clock.Now())
+	if err != nil {
+		return uc.fail(ctx, jobID)
+	}
+	sum, size, err := uc.blobs.Put(ctx, final.StorageKey, bytes.NewReader(output))
+	if err != nil {
+		return uc.fail(ctx, jobID)
+	}
+	final.SetContent(sum, size)
+	if err := uc.tx.WithinTx(ctx, func(ctx context.Context) error {
+		if err := uc.artifacts.Insert(ctx, final); err != nil {
+			return err
+		}
+		return uc.jobs.CompleteWithResult(ctx, jobID, final.ID, uc.clock.Now())
+	}); err != nil {
+		_ = uc.blobs.Delete(ctx, final.StorageKey)
+		return err
+	}
+	return nil
+}
+
+func (uc *ReduceJob) fail(ctx context.Context, jobID uuid.UUID) error {
+	// The public state carries a stable sanitized failure, never parser/storage
+	// internals that may include local paths or implementation details.
+	return uc.jobs.FailReduction(ctx, jobID, "reducer_failed", "final result reduction failed", uc.clock.Now())
+}
+
+func closeAll(closers []io.Closer) {
+	for _, closer := range closers {
+		_ = closer.Close()
+	}
+}
+
+type GetJobResult struct {
+	jobs     JobRepository
+	download *DownloadArtifact
+}
+
+func NewGetJobResult(jobs JobRepository, download *DownloadArtifact) *GetJobResult {
+	return &GetJobResult{jobs: jobs, download: download}
+}
+
+func (uc *GetJobResult) Execute(ctx context.Context, jobID uuid.UUID) (*domain.Artifact, io.ReadCloser, error) {
+	job, err := uc.jobs.Get(ctx, jobID)
+	if err != nil {
+		return nil, nil, err
+	}
+	if job.Status != domain.JobCompleted || job.ResultArtifactID == nil {
+		return nil, nil, domain.ErrArtifactNotFound
+	}
+	return uc.download.Execute(ctx, *job.ResultArtifactID)
+}

coordinator/internal/usecase/usecase_test.go

@@ -55,6 +55,8 @@ type harness struct {
 	getInput    *usecase.GetTaskInput
 	expire      *usecase.ExpireLeases
 	cancel      *usecase.CancelJob
+	reduce      *usecase.ReduceJob
+	jobResult   *usecase.GetJobResult
 }
 
 func newHarness() *harness {
@@ -81,9 +83,88 @@ func newHarness() *harness {
 	h.getInput = usecase.NewGetTaskInput(h.tasks, h.arts, h.blobs)
 	h.expire = usecase.NewExpireLeases(h.tasks, h.jobs, tx, h.clk)
 	h.cancel = usecase.NewCancelJob(h.jobs, h.tasks, tx, h.clk)
+	h.reduce = usecase.NewReduceJob(h.jobs, h.tasks, h.arts, h.blobs, tx, h.clk)
+	h.jobResult = usecase.NewGetJobResult(h.jobs, h.downloadArt)
 	return h
 }
 
+func TestSimilaritySearchReductionCreatesFinalArtifact(t *testing.T) {
+	h := newHarness()
+	jobID := h.seedJob(t, "similarity-search", 2)
+	_, err := h.jobs.Get(ctx, jobID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := h.jobs.UpdateStatus(ctx, jobID, domain.JobRunning, nil); err != nil {
+		t.Fatal(err)
+	}
+	partials := []string{
+		"rank,chembl_id,canonical_smiles,similarity\n1,B,CCC,0.50000048\n",
+		"rank,chembl_id,canonical_smiles,similarity\n1,A,CC,0.50000049\n",
+	}
+	for _, partial := range partials {
+		taskID, attempt := h.leaseOne(t, "w1", "similarity-search")
+		art, err := h.uploadArt.Execute(ctx, usecase.UploadArtifactInput{TaskID: taskID, WorkerID: "w1", Attempt: attempt, Filename: "partial.csv", ContentType: "text/csv", Body: strings.NewReader(partial)})
+		if err != nil {
+			t.Fatal(err)
+		}
+		if _, err := h.complete.Execute(ctx, usecase.CompleteTaskInput{TaskID: taskID, WorkerID: "w1", Attempt: attempt, ResultArtifactID: art.ID}); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if err := h.reduce.Execute(ctx, jobID); err != nil {
+		t.Fatal(err)
+	}
+	progress, err := h.status.Execute(ctx, jobID)
+	if err != nil || progress.Job.Status != domain.JobCompleted {
+		t.Fatalf("status=%s err=%v", progress.Job.Status, err)
+	}
+	art, body, err := h.jobResult.Execute(ctx, jobID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer body.Close()
+	bytes, _ := io.ReadAll(body)
+	if art.Kind != domain.ArtifactFinalResult || string(bytes) != "rank,chembl_id,canonical_smiles,similarity\n1,A,CC,0.500000\n2,B,CCC,0.500000\n" {
+		t.Fatalf("unexpected final %q", bytes)
+	}
+}
+
+func TestSimilaritySearchReductionFailureIsSanitized(t *testing.T) {
+	h := newHarness()
+	jobID := h.seedJob(t, "similarity-search", 1)
+	if err := h.jobs.UpdateStatus(ctx, jobID, domain.JobRunning, nil); err != nil {
+		t.Fatal(err)
+	}
+	taskID, attempt := h.leaseOne(t, "w1", "similarity-search")
+	art, err := h.uploadArt.Execute(ctx, usecase.UploadArtifactInput{
+		TaskID: taskID, WorkerID: "w1", Attempt: attempt, Filename: "partial.csv",
+		ContentType: "text/csv", Body: strings.NewReader("rank,chembl_id,canonical_smiles,similarity\n2,A,CC,0.9\n"),
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := h.complete.Execute(ctx, usecase.CompleteTaskInput{
+		TaskID: taskID, WorkerID: "w1", Attempt: attempt, ResultArtifactID: art.ID,
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if err := h.reduce.Execute(ctx, jobID); err != nil {
+		t.Fatal(err)
+	}
+	job, err := h.jobs.Get(ctx, jobID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if job.Status != domain.JobFailed || job.ErrorCode == nil || *job.ErrorCode != "reducer_failed" ||
+		job.ErrorMessage == nil || *job.ErrorMessage != "final result reduction failed" {
+		t.Fatalf("unexpected failed job: %+v", job)
+	}
+	if job.ResultArtifactID != nil {
+		t.Fatal("failed reduction must not expose a final result")
+	}
+}
+
 // seedJob creates a URI-chunked job with n chunks and returns its id.
 func (h *harness) seedJob(t *testing.T, workload string, n int) uuid.UUID {
 	t.Helper()

coordinator/migrations/0010_job_reduction.down.sql

@@ -0,0 +1,7 @@
+BEGIN;
+
+ALTER TABLE jobs DROP COLUMN IF EXISTS error_message;
+ALTER TABLE jobs DROP COLUMN IF EXISTS error_code;
+ALTER TABLE jobs DROP COLUMN IF EXISTS reducer_started_at;
+
+COMMIT;

coordinator/migrations/0010_job_reduction.up.sql

@@ -0,0 +1,7 @@
+-- PostgreSQL enum values must be committed before they are used by a later
+-- transaction, so this migration intentionally has no BEGIN/COMMIT wrapper.
+ALTER TYPE job_status ADD VALUE IF NOT EXISTS 'reducing';
+
+ALTER TABLE jobs ADD COLUMN IF NOT EXISTS error_code text;
+ALTER TABLE jobs ADD COLUMN IF NOT EXISTS error_message text;
+ALTER TABLE jobs ADD COLUMN IF NOT EXISTS reducer_started_at timestamptz;

docs/api-contract.md

@@ -26,7 +26,8 @@ must be updated in the same change as any behaviour it describes.
 | `POST /tasks/{id}/heartbeat` | renew lease | ✅ done |
 | `POST /tasks/{id}/result` | complete | ✅ done, references `artifact_id` |
 | `POST /tasks/{id}/failure` | fail | ✅ done |
-| `GET /jobs/{id}` | progress | ✅ done |
+| `GET /jobs/{id}` | progress and final-result URI | ✅ done |
+| `GET /jobs/{id}/result` | download final CSV | ✅ done |
 | `PUT /tasks/{id}/artifacts/{name}` | upload partial | ✅ done |
 | `GET /artifacts/{id}/download` | download by id | ✅ done |
 | `POST /jobs/upload` | upload dataset, coordinator chunks it | ✅ done |
@@ -67,6 +68,35 @@ artifact. The coordinator splits the selected TSV rows into shard artifacts
 Each resulting task's claim response carries `input.uri = /tasks/{id}/input`,
 served by §5.4.
 
+## Job progress and final result
+
+```http
+GET /jobs/{job_id}
+Authorization: Bearer <token>
+```
+
+The response contains task counters and a derived status. A successful
+`similarity-search` enters `reducing` after the last shard completes, then
+becomes `completed` only after the coordinator stores its deterministic final
+CSV. At that point `result_uri` is present:
+
+```json
+{
+  "id": "uuid",
+  "status": "completed",
+  "total": 3,
+  "pending": 0,
+  "leased": 0,
+  "completed": 3,
+  "failed": 0,
+  "cancelled": 0,
+  "result_uri": "/jobs/uuid/result"
+}
+```
+
+`GET /jobs/{job_id}/result` downloads that final coordinator-owned CSV. Before
+the job is completed (or for a failed/cancelled job), it returns `404`.
+
 ## Stop a job
 
 ```http

docs/building-workers.md

@@ -56,9 +56,10 @@ Response: `{ "worker_id": "<uuid>", "heartbeat_interval_seconds": 15 }`.
 - **Keep `worker_id`**. Use it as your identity in every later call. Using the
   registered UUID is what lets the coordinator track your liveness (it marks
   workers offline after they go silent).
-- Current diagnostic uploads use `similarity-search` with `query_smiles`. The
-  reference worker accepts the legacy `similarity_search` spelling too. Do not
-  advertise `similarity-graph` until CTX-10 implements cross-shard pair planning.
+- Current distributed uploads use `similarity-search` with `query_smiles`; the
+  coordinator merges completed shard candidates into a final CSV. The reference
+  worker accepts the legacy `similarity_search` spelling too. Do not advertise
+  `similarity-graph` until CTX-10 implements cross-shard pair planning.
 
 ## 2. Claim a task
 

docs/ctx-07-distributed-workload-protocol.md

@@ -4,9 +4,11 @@
 
 This document is the implementation contract for CTX-07. Its generic protocol,
 registry, strict JSON models, and deterministic reduction ordering are
-implemented in `scimesh/distributed/`. It does not implement a molecular
-planner, reducer, API endpoint, database migration, or final artifact. Until
-CTX-08 and CTX-09 are complete, shard CSVs remain diagnostic partial results.
+implemented in `scimesh/distributed/`. CTX-08 implements the molecular
+similarity-search planner, worker adapter, and pure reducer on top of it. This
+document does not implement a coordinator API endpoint, database migration, or
+durable final artifact. Until CTX-09 is complete, shard CSVs remain diagnostic
+partial results.
 
 The protocol gives local scientific workloads a coordinator-independent way to
 validate a job, plan artifact-backed tasks, and later reduce completed outputs.
@@ -154,7 +156,10 @@ rank,chembl_id,canonical_smiles,similarity
 ```
 
 - `rank` is one-based local rank.
-- `similarity` uses the local CLI's six-decimal formatting.
+- `similarity` uses a round-trip decimal representation of the computed float
+  (for Python, `repr(similarity)`). This preserves exact cross-shard ranking;
+  the reducer writes the user-facing final CSV with the local CLI's six-decimal
+  display formatting.
 - Rows are sorted by `(-similarity, chembl_id, canonical_smiles)` for
   `threshold_direction=greater`, or `(similarity, chembl_id,
   canonical_smiles)` for `less`.
@@ -203,7 +208,7 @@ multiplicity. Reduction is independent of worker completion order and uses
 
 ## Deferred work
 
-CTX-08 implements the similarity-search planner, runner adapter, reducer, and
-comparison against the local CLI. CTX-09 persists the final artifact and job
-state. CTX-10 defines graph-specific triangular block plans; it must not reuse
-the search shard scheme without its pair-coverage invariants.
+CTX-09 materializes the planned shard files as coordinator artifacts, invokes
+the registered reducer once, and persists its final artifact/job state. CTX-10
+defines graph-specific triangular block plans; it must not reuse the search
+shard scheme without its pair-coverage invariants.

docs/openapi.yaml

@@ -97,8 +97,9 @@ paths:
         multipart/form-data. The text fields (`workload`, `parameters`,
         `chunk_rows`, `max_rows`) MUST precede the `file` part: the file is streamed, not
         buffered, so the fields have to be parsed before it arrives. Currently
-        only diagnostic `similarity-search` with `parameters.query_smiles` is
-        accepted; distributed graph planning is not implemented.
+        only `similarity-search` with `parameters.query_smiles` is accepted.
+        When every shard succeeds, the coordinator merges their candidates into
+        one final CSV; distributed graph planning is not implemented.
       requestBody:
         required: true
         content:
@@ -124,7 +125,9 @@ paths:
         - $ref: "#/components/parameters/JobID"
       responses:
         "200":
-          description: Progress counts and derived status.
+          description: >
+            Progress counts and derived status. A completed similarity-search
+            response includes `result_uri` for its final CSV.
           content:
             application/json:
               schema: { $ref: "#/components/schemas/JobProgress" }
@@ -132,6 +135,21 @@ paths:
         "401": { $ref: "#/components/responses/Unauthorized" }
         "404": { $ref: "#/components/responses/NotFound" }
 
+  /jobs/{job_id}/result:
+    get:
+      tags: [jobs]
+      summary: Download a completed job's final result
+      parameters:
+        - $ref: "#/components/parameters/JobID"
+      responses:
+        "200":
+          description: Final coordinator-owned CSV.
+          content:
+            text/csv:
+              schema: { type: string, format: binary }
+        "401": { $ref: "#/components/responses/Unauthorized" }
+        "404": { $ref: "#/components/responses/NotFound" }
+
   /jobs/{job_id}/cancel:
     post:
       tags: [jobs]
@@ -483,6 +501,13 @@ components:
         completed: { type: integer }
         failed: { type: integer }
         cancelled: { type: integer }
+        result_uri:
+          type: string
+          description: Present only when the final result is available.
+          example: /jobs/5a4c3a7f-ccfc-47d6-b78d-2d1fa565bafd/result
+        error_code:
+          type: string
+          description: Sanitized terminal reducer failure code, when applicable.
 
     ClaimRequest:
       type: object
@@ -577,7 +602,7 @@ components:
 
     JobStatus:
       type: string
-      enum: [pending, running, completed, failed, cancelled]
+      enum: [pending, running, reducing, completed, failed, cancelled]
 
     TaskStatus:
       type: string

docs/worker-daemon-task.md

@@ -122,7 +122,10 @@ Content-Type: application/json
   },
   "metrics": {
     "elapsed_seconds": 12.4,
-    "processed_rows": 10000
+    "scanned_rows": 10000,
+    "valid_molecules": 9876,
+    "invalid_smiles": 124,
+    "matches_emitted": 20
   }
 }
 ```
@@ -189,8 +192,11 @@ class Runner(Protocol):
         """Run one task and return output artifacts plus safe metrics."""
 ```
 
-`SciMeshRunner` should map `workload` and validated parameters to the existing
-SciMesh CLI. For example, a `similarity-search` task invokes:
+`SciMeshRunner` maps an allowlisted workload and validated parameters to the
+local SciMesh reference functions. A planned `similarity-search` task contains
+a resolved `query_smiles` (never `query_id`) and writes one exact local top-k
+partial CSV plus the metrics above. Legacy single-shard tasks may still use the
+CLI compatibility path:
 
 ```text
 scimesh similarity-search <local-input> --query-id ... --output <task-dir>/result.csv

scimesh/distributed/__init__.py

@@ -12,17 +12,25 @@ from .models import (
     FinalResult,
     PlannedTask,
 )
-from .registry import DistributedWorkloadRegistry, PlanningService, WorkloadDescription
+from .registry import (
+    DistributedWorkloadRegistry,
+    PlanningService,
+    WorkloadDescription,
+    default_distributed_registry,
+)
+from .similarity_search import SimilaritySearchDistributedWorkload
 from .workload import DistributedWorkload
 
 __all__ = [
     "ArtifactReference",
     "CompletedPartial",
+    "default_distributed_registry",
     "DistributedPlan",
     "DistributedWorkload",
     "DistributedWorkloadRegistry",
     "FinalResult",
     "PlannedTask",
     "PlanningService",
+    "SimilaritySearchDistributedWorkload",
     "WorkloadDescription",
 ]

scimesh/distributed/registry.py

@@ -50,7 +50,8 @@ class PlanningService:
 
     It writes neither jobs nor artifacts.  A Go coordinator bridge can therefore
     validate and produce a plan before opening its own all-or-nothing persistence
-    transaction; CTX-08/09 will implement that concrete bridge and reducers.
+    transaction; CTX-09 will implement that concrete bridge and durable result
+    orchestration.
     """
 
     def __init__(self, registry: DistributedWorkloadRegistry) -> None:
@@ -94,3 +95,14 @@ class PlanningService:
         if not isinstance(result, FinalResult):
             raise ValueError("distributed reducer must return a FinalResult")
         return result
+
+
+def default_distributed_registry() -> DistributedWorkloadRegistry:
+    """Return the currently supported distributed scientific workloads."""
+    # Delayed import keeps the generic registry independent of concrete RDKit
+    # workloads and avoids making the contract layer import application setup.
+    from .similarity_search import SimilaritySearchDistributedWorkload
+
+    registry = DistributedWorkloadRegistry()
+    registry.register(SimilaritySearchDistributedWorkload())
+    return registry

scimesh/distributed/similarity_search.py

@@ -0,0 +1,380 @@
+"""Distributed planning and reduction for exact molecular similarity search."""
+
+from __future__ import annotations
+
+import csv
+import hashlib
+import heapq
+import math
+from pathlib import Path
+from typing import Any, Iterator, Mapping, Sequence
+from uuid import UUID, uuid5
+
+from rdkit import Chem
+
+from scimesh.chemistry.dataset import MoleculeRecord, find_molecule_by_id, parse_smiles
+from scimesh.chemistry.fingerprints import FP_RADIUS, FP_SIZE
+from scimesh.workloads.similarity_search import (
+    SimilarityMatch,
+    _HeapEntry,
+    search_similar,
+    write_search_results,
+)
+
+from .models import ArtifactReference, CompletedPartial, DistributedPlan, FinalResult, PlannedTask
+
+
+_TSV_CONTENT_TYPE = "text/tab-separated-values"
+_CSV_CONTENT_TYPE = "text/csv"
+_SEARCH_COLUMNS = ("rank", "chembl_id", "canonical_smiles", "similarity")
+_REQUIRED_COLUMNS = {"chembl_id", "canonical_smiles"}
+
+
+def write_similarity_search_partial(output_path: Path, matches: Sequence[SimilarityMatch]) -> None:
+    """Write a worker partial with a round-trip score, not display rounding.
+
+    The public final CSV continues to use the local CLI's six-decimal display.
+    A reducer needs the full binary float representation to rank candidates
+    from separate shards exactly as the single-process reference does.
+    """
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    with output_path.open("w", encoding="utf-8", newline="") as destination:
+        writer = csv.DictWriter(destination, fieldnames=_SEARCH_COLUMNS)
+        writer.writeheader()
+        for rank, match in enumerate(matches, start=1):
+            writer.writerow({
+                "rank": rank,
+                "chembl_id": match.molecule_id,
+                "canonical_smiles": match.smiles,
+                "similarity": repr(match.similarity),
+            })
+
+
+class SimilaritySearchDistributedWorkload:
+    """Planner/reducer for exact global top-k Tanimoto similarity search."""
+
+    name = "similarity-search"
+    description = "Exact top-k molecular similarity search over deterministic TSV shards."
+
+    def validate_job(self, parameters: Mapping[str, object]) -> None:
+        allowed = {
+            "query_id", "query_smiles", "top_k", "threshold",
+            "threshold_direction", "max_rows", "progress_every",
+        }
+        unknown = set(parameters) - allowed
+        if unknown:
+            raise ValueError(f"unsupported similarity-search parameters: {', '.join(sorted(unknown))}")
+        query_id = parameters.get("query_id")
+        query_smiles = parameters.get("query_smiles")
+        if (query_id is None) == (query_smiles is None):
+            raise ValueError("exactly one of query_id or query_smiles is required")
+        if query_id is not None:
+            self._string(query_id, "query_id")
+        if query_smiles is not None:
+            self._string(query_smiles, "query_smiles")
+        self._positive_int(parameters.get("top_k", 20), "top_k")
+        if "max_rows" in parameters:
+            self._positive_int(parameters["max_rows"], "max_rows")
+        if "progress_every" in parameters:
+            self._nonnegative_int(parameters["progress_every"], "progress_every")
+        if "threshold" in parameters:
+            self._unit_interval(parameters["threshold"], "threshold")
+        if "threshold_direction" in parameters and parameters["threshold_direction"] not in {"greater", "less"}:
+            raise ValueError("threshold_direction must be 'greater' or 'less'")
+
+    def plan(
+        self,
+        input_path: Path,
+        input_artifact_id: str,
+        parameters: Mapping[str, object],
+        shard_rows: int,
+        workspace: Path,
+    ) -> DistributedPlan:
+        self.validate_job(parameters)
+        if not input_path.is_file():
+            raise ValueError("input_path must be a readable dataset file")
+        if isinstance(shard_rows, bool) or not isinstance(shard_rows, int) or shard_rows < 1:
+            raise ValueError("shard_rows must be a positive integer")
+        try:
+            input_id = UUID(input_artifact_id)
+        except ValueError as error:
+            raise ValueError("input_artifact_id must be a UUID") from error
+
+        query_smiles, query_source = self._resolve_query(input_path, parameters)
+        resolved = self._resolved_parameters(parameters, query_smiles, query_source)
+        workspace.mkdir(parents=True, exist_ok=True)
+        shard_paths: list[Path] = []
+        try:
+            shard_paths = self._write_shards(input_path, workspace, shard_rows, resolved.get("max_rows"))
+            tasks = tuple(
+                PlannedTask(
+                    chunk_index=index,
+                    input_artifact=ArtifactReference(
+                        artifact_id=str(uuid5(input_id, f"scimesh:similarity-search:shard:{index}")),
+                        sha256=_sha256_file(path),
+                        content_type=_TSV_CONTENT_TYPE,
+                    ),
+                    parameters=self._task_parameters(resolved),
+                )
+                for index, path in enumerate(shard_paths)
+            )
+        except Exception:
+            for path in shard_paths:
+                path.unlink(missing_ok=True)
+            raise
+        return DistributedPlan(self.name, resolved, tasks)
+
+    def reduce(
+        self,
+        partial_results: Sequence[CompletedPartial],
+        parameters: Mapping[str, object],
+        workspace: Path,
+    ) -> FinalResult:
+        """Merge materialized partial CSVs into one deterministic final CSV.
+
+        The coordinator bridge materializes each downloaded artifact at
+        ``workspace / artifact_id`` before it calls this method. Those local
+        paths are an ephemeral bridge detail, never present in the plan or task
+        payload. CTX-09 owns the durable final-artifact upload and job state.
+        """
+        if not partial_results:
+            raise ValueError("at least one partial result is required")
+        resolved = self._validate_resolved_parameters(parameters)
+        top_k = resolved["top_k"]
+        direction = resolved["threshold_direction"]
+        heap: list[_HeapEntry] = []
+
+        ordered_partials = tuple(sorted(partial_results, key=lambda partial: partial.chunk_index))
+        indexes = [partial.chunk_index for partial in ordered_partials]
+        if len(indexes) != len(set(indexes)):
+            raise ValueError("partial results must have unique chunk_index values")
+        for partial in ordered_partials:
+            path = workspace / partial.artifact.artifact_id
+            if not path.is_file():
+                raise ValueError("materialized partial result is missing")
+            if _sha256_file(path) != partial.artifact.sha256:
+                raise ValueError("materialized partial result checksum does not match its artifact reference")
+            for match in self._read_partial(path, direction):
+                rank_key = match.sort_key(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 = sorted((entry.match for entry in heap), key=lambda match: match.sort_key(direction))
+        output = workspace / "result.csv"
+        write_search_results(output, matches)
+        final_id = uuid5(
+            UUID(ordered_partials[0].artifact.artifact_id),
+            "scimesh:similarity-search:final:" + ",".join(
+                partial.artifact.artifact_id for partial in ordered_partials
+            ),
+        )
+        return FinalResult(
+            ArtifactReference(str(final_id), _sha256_file(output), _CSV_CONTENT_TYPE),
+            {"matches_emitted": len(matches), "partial_count": len(ordered_partials)},
+        )
+
+    def _resolve_query(
+        self, input_path: Path, parameters: Mapping[str, object]
+    ) -> tuple[str, dict[str, str]]:
+        query_id = parameters.get("query_id")
+        if isinstance(query_id, str):
+            record = find_molecule_by_id(input_path, query_id)
+            return Chem.MolToSmiles(record.molecule, canonical=True), {"kind": "chembl_id", "value": query_id}
+        supplied = parameters["query_smiles"]
+        assert isinstance(supplied, str)  # checked by validate_job
+        molecule = parse_smiles(supplied)
+        if molecule is None:
+            raise ValueError("query_smiles is invalid")
+        return Chem.MolToSmiles(molecule, canonical=True), {"kind": "smiles", "value": supplied}
+
+    def _resolved_parameters(
+        self, parameters: Mapping[str, object], query_smiles: str, query_source: Mapping[str, str]
+    ) -> dict[str, object]:
+        resolved: dict[str, object] = {
+            "query_smiles": query_smiles,
+            "query_source": dict(query_source),
+            "top_k": self._positive_int(parameters.get("top_k", 20), "top_k"),
+            "threshold_direction": parameters.get("threshold_direction", "greater"),
+            "fingerprint": {"algorithm": "morgan", "radius": FP_RADIUS, "fp_size": FP_SIZE},
+        }
+        if "threshold" in parameters:
+            resolved["threshold"] = self._unit_interval(parameters["threshold"], "threshold")
+        if "max_rows" in parameters:
+            resolved["max_rows"] = self._positive_int(parameters["max_rows"], "max_rows")
+        if "progress_every" in parameters:
+            resolved["progress_every"] = self._nonnegative_int(parameters["progress_every"], "progress_every")
+        return resolved
+
+    def _validate_resolved_parameters(self, parameters: Mapping[str, object]) -> dict[str, object]:
+        query_smiles = self._string(parameters.get("query_smiles"), "query_smiles")
+        if parse_smiles(query_smiles) is None:
+            raise ValueError("query_smiles is invalid")
+        resolved = self._resolved_parameters(
+            parameters,
+            Chem.MolToSmiles(parse_smiles(query_smiles), canonical=True),
+            {"kind": "resolved", "value": query_smiles},
+        )
+        # A reducer receives immutable plan metadata, whose query source and
+        # fixed fingerprint are observational context rather than worker input.
+        if "fingerprint" in parameters:
+            fingerprint = parameters["fingerprint"]
+            if fingerprint != {"algorithm": "morgan", "radius": FP_RADIUS, "fp_size": FP_SIZE}:
+                raise ValueError("resolved fingerprint does not match SciMesh defaults")
+        return resolved
+
+    @staticmethod
+    def _task_parameters(resolved: Mapping[str, object]) -> dict[str, object]:
+        # max_rows is applied before sharding. Passing it to each task would
+        # silently scan N rows per shard instead of the requested global prefix.
+        return {
+            key: value for key, value in resolved.items()
+            if key in {"query_smiles", "top_k", "threshold", "threshold_direction", "progress_every"}
+        }
+
+    def _write_shards(
+        self, input_path: Path, workspace: Path, shard_rows: int, max_rows: object
+    ) -> list[Path]:
+        limit = int(max_rows) if isinstance(max_rows, int) else None
+        paths: list[Path] = []
+        current: Path | None = None
+        destination = None
+        rows_in_shard = 0
+        seen_rows = 0
+        try:
+            with input_path.open("r", encoding="utf-8", newline="") as source:
+                reader = csv.DictReader(source, delimiter="\t")
+                fieldnames = reader.fieldnames or []
+                if not _REQUIRED_COLUMNS.issubset(set(fieldnames)):
+                    missing = sorted(_REQUIRED_COLUMNS - set(fieldnames))
+                    raise ValueError(f"dataset is missing required columns: {', '.join(missing)}")
+                for row in reader:
+                    if limit is not None and seen_rows >= limit:
+                        break
+                    if destination is None or rows_in_shard == shard_rows:
+                        if destination is not None:
+                            destination.close()
+                        current = workspace / f"shard-{len(paths)}.tsv"
+                        destination = current.open("w", encoding="utf-8", newline="")
+                        writer = csv.DictWriter(destination, fieldnames=fieldnames, delimiter="\t", lineterminator="\n")
+                        writer.writeheader()
+                        paths.append(current)
+                        rows_in_shard = 0
+                    writer.writerow(row)
+                    rows_in_shard += 1
+                    seen_rows += 1
+        finally:
+            if destination is not None:
+                destination.close()
+        if not paths:
+            raise ValueError("dataset has no data rows")
+        return paths
+
+    @staticmethod
+    def _read_partial(path: Path, direction: object) -> Iterator[SimilarityMatch]:
+        if not path.is_file():
+            raise ValueError("materialized partial result is missing")
+        if direction not in {"greater", "less"}:
+            raise ValueError("threshold_direction must be 'greater' or 'less'")
+        previous_key: tuple[float, str, str] | None = None
+        with path.open("r", encoding="utf-8", newline="") as source:
+            reader = csv.DictReader(source)
+            if tuple(reader.fieldnames or ()) != _SEARCH_COLUMNS:
+                raise ValueError("partial result has an invalid CSV header")
+            for expected_rank, row in enumerate(reader, start=1):
+                if set(row) != set(_SEARCH_COLUMNS) or row["rank"] != str(expected_rank):
+                    raise ValueError("partial result has an invalid rank")
+                try:
+                    similarity = float(row["similarity"])
+                except (TypeError, ValueError) as error:
+                    raise ValueError("partial result has an invalid similarity") from error
+                if not math.isfinite(similarity) or not 0 <= similarity <= 1:
+                    raise ValueError("partial result has an invalid similarity")
+                match = SimilarityMatch(similarity, row["chembl_id"], row["canonical_smiles"])
+                key = match.sort_key(direction)
+                if previous_key is not None and key < previous_key:
+                    raise ValueError("partial result is not sorted deterministically")
+                previous_key = key
+                yield match
+
+    @staticmethod
+    def _string(value: object, name: str) -> str:
+        if not isinstance(value, str) or not value.strip() or len(value) > 200:
+            raise ValueError(f"{name} must be a non-empty string")
+        return value
+
+    @staticmethod
+    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
+
+    @staticmethod
+    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
+
+    @staticmethod
+    def _unit_interval(value: object, name: str) -> float:
+        if isinstance(value, bool) or not isinstance(value, (int, float)) or not math.isfinite(value) or not 0 <= value <= 1:
+            raise ValueError(f"{name} must be a number between 0 and 1")
+        return float(value)
+
+
+def run_similarity_search_shard(
+    input_path: Path, parameters: Mapping[str, object], output_path: Path
+) -> dict[str, int]:
+    """Run one planned shard using the local reference implementation.
+
+    This is the worker adapter used by CTX-08. It deliberately accepts only
+    resolved ``query_smiles``: resolving an identifier independently in each
+    shard would make the distributed search scientifically invalid.
+    """
+    allowed = {"query_smiles", "top_k", "threshold", "threshold_direction", "progress_every"}
+    unknown = set(parameters) - allowed
+    if unknown:
+        raise ValueError(f"unsupported similarity-search parameters: {', '.join(sorted(unknown))}")
+    query_smiles = parameters.get("query_smiles")
+    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 = SimilaritySearchDistributedWorkload._positive_int(parameters.get("top_k", 20), "top_k")
+    threshold = None
+    if "threshold" in parameters:
+        threshold = SimilaritySearchDistributedWorkload._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'")
+    progress_every = 0
+    if "progress_every" in parameters:
+        progress_every = SimilaritySearchDistributedWorkload._nonnegative_int(
+            parameters["progress_every"], "progress_every"
+        )
+    result = search_similar(
+        input_path,
+        MoleculeRecord("query", query_smiles, molecule),
+        top_k=top_k,
+        progress_every=progress_every,
+        threshold=threshold,
+        threshold_direction=direction,
+    )
+    write_similarity_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 _sha256_file(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as source:
+        for block in iter(lambda: source.read(1024 * 1024), b""):
+            digest.update(block)
+    return digest.hexdigest()

scimesh/worker/daemon.py

@@ -9,6 +9,7 @@ from pathlib import Path
 import random
 import re
 import shutil
+import subprocess
 import threading
 import time
 from datetime import datetime, timezone
@@ -202,12 +203,23 @@ class WorkerDaemon:
     def _report_failure(self, task: ClaimedTask, error: Exception) -> None:
         message = self._sanitize_error_message(error)
         try:
-            self.coordinator.fail(task, {"worker_id": self._worker_id(), "attempt": task.attempt, "error_code": type(error).__name__, "error_message": message})
+            self.coordinator.fail(task, {
+                "worker_id": self._worker_id(),
+                "attempt": task.attempt,
+                "error_code": type(error).__name__,
+                "error_message": message,
+                "retryable": self._is_retryable(error),
+            })
         except CoordinatorTransientError:
             raise
         except Exception:
             self._log("failed", task, error_type="FailureReportError")
 
+    @staticmethod
+    def _is_retryable(error: Exception) -> bool:
+        """Retry transient worker/transport failures, never invalid scientific input."""
+        return not isinstance(error, (ValueError, FileNotFoundError, subprocess.CalledProcessError))
+
     def _sanitize_error_message(self, error: Exception) -> str:
         """Keep coordinator-visible failures useful without exposing local paths."""
         message = str(error).replace(str(self.config.work_dir), "<worker-dir>")

scimesh/worker/runners.py

@@ -7,6 +7,8 @@ import subprocess
 import sys
 from typing import Protocol
 
+from scimesh.distributed.similarity_search import run_similarity_search_shard
+
 from .models import ClaimedTask, ProducedArtifact, RunResult
 
 
@@ -34,6 +36,12 @@ class SciMeshRunner:
             query_id, query_smiles = params.get("query_id"), params.get("query_smiles")
             if (query_id is None) == (query_smiles is None):
                 raise ValueError("exactly one of query_id or query_smiles is required")
+            if query_smiles is not None and "max_rows" not in params:
+                metrics = run_similarity_search_shard(input_path, params, output_path)
+                return RunResult((ProducedArtifact(output_path, "text/csv"),), metrics)
+            # Legacy URI jobs may still use query_id or an explicitly task-local
+            # max_rows value. CTX-08 plans never create those payloads; retain
+            # CLI execution only for backwards compatibility at this boundary.
             top_k = self._positive_int(params, "top_k", default=20)
             command += ["--query-id", self._string(params, "query_id")] if query_id is not None else ["--query-smiles", self._string(params, "query_smiles")]
             command += ["--top-k", str(top_k)]

tests/test_distributed_similarity_search.py

@@ -0,0 +1,184 @@
+"""Scientific reference tests for the CTX-08 distributed search workload."""
+
+from __future__ import annotations
+
+import csv
+import hashlib
+from pathlib import Path
+from uuid import NAMESPACE_URL, uuid5
+
+import pytest
+
+from scimesh.chemistry.dataset import find_molecule_by_id
+from scimesh.distributed import (
+    ArtifactReference,
+    CompletedPartial,
+    PlanningService,
+    default_distributed_registry,
+)
+from scimesh.distributed.registry import DistributedWorkloadRegistry
+from scimesh.distributed.similarity_search import (
+    SimilaritySearchDistributedWorkload,
+    run_similarity_search_shard,
+    write_similarity_search_partial,
+)
+from scimesh.workloads.similarity_search import search_similar, write_search_results
+
+
+def make_dataset(path: Path) -> None:
+    path.write_text(
+        "chembl_id\tcanonical_smiles\textra\n"
+        "CHEMBL_QUERY\tCCO\tquery\n"
+        "CHEMBL_A\tCCCO\ta\n"
+        "CHEMBL_B\tCCCC\tb\n"
+        "CHEMBL_INVALID\tnot-a-smiles\tbad\n"
+        "CHEMBL_DUPLICATE\tCCO\tduplicate\n"
+        "CHEMBL_C\tCCN\tc\n",
+        encoding="utf-8",
+    )
+
+
+def planner() -> tuple[PlanningService, SimilaritySearchDistributedWorkload]:
+    workload = SimilaritySearchDistributedWorkload()
+    registry = DistributedWorkloadRegistry()
+    registry.register(workload)
+    return PlanningService(registry), workload
+
+
+def test_default_registry_exposes_only_supported_distributed_search() -> None:
+    assert [item.name for item in default_distributed_registry().descriptions()] == ["similarity-search"]
+
+
+def checksum(path: Path) -> str:
+    return hashlib.sha256(path.read_bytes()).hexdigest()
+
+
+def test_query_id_is_resolved_once_before_deterministic_shards(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    dataset = tmp_path / "chembl.tsv"
+    workspace = tmp_path / "workspace"
+    make_dataset(dataset)
+    service, _ = planner()
+    calls = 0
+    real_find = find_molecule_by_id
+
+    def count_find(path: Path, query_id: str) -> MoleculeRecord:
+        nonlocal calls
+        calls += 1
+        return real_find(path, query_id)
+
+    monkeypatch.setattr("scimesh.distributed.similarity_search.find_molecule_by_id", count_find)
+    input_id = str(uuid5(NAMESPACE_URL, "dataset"))
+    plan = service.plan(
+        "similarity-search", dataset, input_id,
+        {"query_id": "CHEMBL_QUERY", "top_k": 3, "max_rows": 5, "progress_every": 0},
+        2, workspace,
+    )
+
+    assert calls == 1
+    assert plan.resolved_parameters["query_smiles"] == "CCO"
+    assert plan.resolved_parameters["query_source"] == {"kind": "chembl_id", "value": "CHEMBL_QUERY"}
+    assert [task.chunk_index for task in plan.tasks] == [0, 1, 2]
+    assert all("query_id" not in task.parameters for task in plan.tasks)
+    assert all("max_rows" not in task.parameters for task in plan.tasks)
+    assert all(task.parameters["query_smiles"] == "CCO" for task in plan.tasks)
+    assert [
+        sum(1 for _ in path.open(encoding="utf-8")) - 1
+        for path in sorted(workspace.glob("shard-*.tsv"))
+    ] == [2, 2, 1]
+
+
+def test_distributed_reduction_matches_single_process_reference(tmp_path: Path) -> None:
+    dataset = tmp_path / "chembl.tsv"
+    workspace = tmp_path / "workspace"
+    make_dataset(dataset)
+    service, workload = planner()
+    plan = service.plan(
+        "similarity-search", dataset, str(uuid5(NAMESPACE_URL, "dataset")),
+        {"query_smiles": "CCO", "top_k": 3, "threshold": 0.0}, 2, workspace,
+    )
+
+    partials: list[CompletedPartial] = []
+    # Worker two finishes the latter shards first. Worker one loses its first
+    # attempt for shard zero, then retries it last. The reducer must remain
+    # independent of both completion and retry order.
+    for task in reversed(plan.tasks):
+        shard = workspace / f"shard-{task.chunk_index}.tsv"
+        temporary_partial = workspace / f"worker-output-{task.chunk_index}.csv"
+        metrics = run_similarity_search_shard(shard, task.parameters, temporary_partial)
+        partial_id = str(uuid5(NAMESPACE_URL, f"partial:{task.chunk_index}"))
+        partials.append(
+            CompletedPartial(
+                task.chunk_index,
+                ArtifactReference(
+                    partial_id, checksum(temporary_partial), "text/csv",
+                ),
+                metrics,
+            )
+        )
+        # The reducer materializes result files under their own coordinator IDs,
+        # not shard input IDs. Keep this fixture faithful to that boundary.
+        temporary_partial.rename(workspace / partial_id)
+
+    final = workload.reduce(tuple(partials), plan.resolved_parameters, workspace)
+    reference = tmp_path / "reference.csv"
+    query_record = find_molecule_by_id(dataset, "CHEMBL_QUERY")
+    write_search_results(reference, search_similar(dataset, query_record, top_k=3, threshold=0.0).matches)
+
+    assert (workspace / "result.csv").read_bytes() == reference.read_bytes()
+    assert final.metrics == {"matches_emitted": 3, "partial_count": 3}
+    rows = list(csv.DictReader((workspace / "result.csv").open(encoding="utf-8")))
+    assert {row["chembl_id"] for row in rows}.isdisjoint({"CHEMBL_QUERY", "CHEMBL_DUPLICATE"})
+
+
+def test_reducer_rejects_unsorted_or_invalid_partial_csv(tmp_path: Path) -> None:
+    workspace = tmp_path / "workspace"
+    workspace.mkdir()
+    workload = SimilaritySearchDistributedWorkload()
+    artifact_id = str(uuid5(NAMESPACE_URL, "bad"))
+    partial_path = workspace / artifact_id
+    partial_path.write_text(
+        "rank,chembl_id,canonical_smiles,similarity\n1,A,CC,0.1\n2,B,CCC,0.9\n",
+        encoding="utf-8",
+    )
+    artifact = ArtifactReference(artifact_id, checksum(partial_path), "text/csv")
+
+    with pytest.raises(ValueError, match="not sorted"):
+        workload.reduce(
+            (CompletedPartial(0, artifact, {"scanned_rows": 2}),),
+            {"query_smiles": "CCO", "top_k": 2, "threshold_direction": "greater", "fingerprint": {"algorithm": "morgan", "radius": 2, "fp_size": 2048}},
+            workspace,
+        )
+
+
+def test_partial_csv_preserves_exact_scores_for_global_ranking(tmp_path: Path) -> None:
+    partial = tmp_path / "partial.csv"
+    # Both values look identical in a six-decimal final CSV. The exact value
+    # must survive shard transport so the global reducer can still rank them.
+    from scimesh.workloads.similarity_search import SimilarityMatch
+
+    write_similarity_search_partial(
+        partial,
+        [SimilarityMatch(0.50000049, "A", "CC"), SimilarityMatch(0.50000048, "B", "CCC")],
+    )
+    values = list(csv.DictReader(partial.open(encoding="utf-8")))
+    assert values[0]["similarity"] == repr(0.50000049)
+    assert values[1]["similarity"] == repr(0.50000048)
+
+
+def test_planner_rejects_fingerprint_override_and_invalid_query(tmp_path: Path) -> None:
+    dataset = tmp_path / "chembl.tsv"
+    make_dataset(dataset)
+    service, _ = planner()
+
+    with pytest.raises(ValueError, match="unsupported similarity-search parameters"):
+        service.plan(
+            "similarity-search", dataset, str(uuid5(NAMESPACE_URL, "dataset")),
+            {"query_smiles": "CCO", "fingerprint": {"radius": 1}}, 2, tmp_path / "workspace",
+        )
+    with pytest.raises(ValueError, match="query_smiles is invalid"):
+        service.plan(
+            "similarity-search", dataset, str(uuid5(NAMESPACE_URL, "dataset")),
+            {"query_smiles": "invalid"}, 2, tmp_path / "workspace",
+        )

tests/test_worker_daemon.py

@@ -106,6 +106,90 @@ def test_claims_runs_uploads_and_submits_csv(tmp_path: Path) -> None:
     }
 
 
+def test_worker_executes_a_resolved_similarity_search_shard(tmp_path: Path) -> None:
+    content = b"chembl_id\tcanonical_smiles\nQUERY\tCCO\nMATCH\tCCCO\nINVALID\tnot-a-smiles\n"
+    task = make_task(content)
+    task = ClaimedTask(
+        task.task_id, task.attempt, task.lease_expires_at, task.workload, task.input,
+        {"query_smiles": "CCO", "top_k": 5, "progress_every": 0},
+    )
+    worker, coordinator, artifacts, _, _ = daemon(tmp_path, task, content)
+    worker.runner = SciMeshRunner()
+
+    assert worker.run_once() == RunOnceOutcome(claimed=True, completed=True)
+    output = artifacts.uploaded[0][2].read_text(encoding="utf-8")
+    assert output.startswith("rank,chembl_id,canonical_smiles,similarity\n")
+    metrics = coordinator.submissions[0]["metrics"]
+    assert metrics["scanned_rows"] == 3
+    assert metrics["valid_molecules"] == 2
+    assert metrics["invalid_smiles"] == 1
+    assert metrics["matches_emitted"] == 1
+    assert isinstance(metrics["elapsed_seconds"], float)
+
+
+def test_two_workers_complete_resolved_shards_after_one_retry(tmp_path: Path) -> None:
+    content = b"chembl_id\tcanonical_smiles\nQUERY\tCCO\nMATCH\tCCCO\n"
+    first = make_task(content)
+    first = ClaimedTask(
+        "retry-task", 1, first.lease_expires_at, "similarity-search", first.input,
+        {"query_smiles": "CCO", "top_k": 5},
+    )
+    second = ClaimedTask(
+        "other-task", 1, first.lease_expires_at, "similarity-search", first.input,
+        {"query_smiles": "CCO", "top_k": 5},
+    )
+
+    class RetryCoordinator(FakeCoordinator):
+        def __init__(self) -> None:
+            super().__init__(None)
+            self.queue = [first, second]
+            self.claimants: list[str] = []
+
+        def claim(self, worker_id: str, capabilities: tuple[str, ...]) -> ClaimedTask | None:
+            self.claimants.append(worker_id)
+            return self.queue.pop(0) if self.queue else None
+
+        def fail(self, task: ClaimedTask, payload: dict) -> None:
+            self.failures.append(payload)
+            if task.task_id == "retry-task" and task.attempt == 1 and payload["retryable"]:
+                self.queue.append(
+                    ClaimedTask(
+                        task.task_id, 2, task.lease_expires_at, task.workload, task.input,
+                        task.parameters,
+                    )
+                )
+
+    class FailFirstAttempt:
+        def __init__(self) -> None:
+            self.calls = 0
+            self.delegate = SciMeshRunner()
+
+        def run(self, task: ClaimedTask, task_dir: Path) -> RunResult:
+            self.calls += 1
+            if self.calls == 1:
+                raise RuntimeError("simulated retryable shard failure")
+            return self.delegate.run(task, task_dir)
+
+    coordinator = RetryCoordinator()
+    artifacts = FakeArtifacts(content)
+    worker_a = WorkerDaemon(
+        WorkerConfig("https://example.test", "worker-a", tmp_path / "worker-a"),
+        coordinator, artifacts, FailFirstAttempt(),
+    )
+    worker_b = WorkerDaemon(
+        WorkerConfig("https://example.test", "worker-b", tmp_path / "worker-b"),
+        coordinator, artifacts, SciMeshRunner(),
+    )
+
+    assert worker_a.run_once() == RunOnceOutcome(claimed=True, completed=False)
+    assert worker_b.run_once() == RunOnceOutcome(claimed=True, completed=True)
+    assert worker_a.run_once() == RunOnceOutcome(claimed=True, completed=True)
+    assert coordinator.claimants == ["worker-a", "worker-b", "worker-a"]
+    assert len(coordinator.failures) == 1
+    assert coordinator.failures[0]["retryable"] is True
+    assert len(coordinator.submissions) == 2
+
+
 def test_no_task_does_not_create_directory(tmp_path: Path) -> None:
     worker, _, _, runner, config = daemon(tmp_path, None, b"")
     assert worker.run_once() == RunOnceOutcome(claimed=False, completed=False)
@@ -161,6 +245,7 @@ def test_interrupting_an_active_task_reports_a_sanitized_failure(tmp_path: Path)
             "attempt": 1,
             "error_code": "InterruptedError",
             "error_message": "worker interrupted by operator",
+            "retryable": True,
         }
     ]
 
@@ -234,6 +319,7 @@ def test_bad_checksum_reports_failure_without_running(tmp_path: Path) -> None:
     assert worker.run_once() == RunOnceOutcome(claimed=True, completed=False)
     assert runner.calls == 0
     assert coordinator.failures[0]["error_code"] == "ValueError"
+    assert coordinator.failures[0]["retryable"] is False
     assert not coordinator.submissions
 
 
@@ -356,30 +442,35 @@ def test_runner_maps_graph_and_smiles_search_parameters(tmp_path: Path, monkeypa
     runner = SciMeshRunner()
     graph = ClaimedTask("graph", 1, "2026-07-30T00:00:00Z", "similarity-graph", InputArtifact("https://example/input", "x"), {"threshold": 0.2, "threshold_direction": "less", "block_size": 42, "max_rows": 7, "progress_every": 0})
     search = ClaimedTask("search", 1, "2026-07-30T00:00:00Z", "similarity-search", InputArtifact("https://example/input", "x"), {"query_smiles": "CCO", "top_k": 3})
+    search_dir = tmp_path / "search"
+    search_dir.mkdir()
+    (search_dir / "input").write_text(
+        "chembl_id\tcanonical_smiles\nA\tCCO\nB\tCCCO\n", encoding="utf-8"
+    )
     runner.run(graph, tmp_path / "graph")
-    runner.run(search, tmp_path / "search")
+    result = runner.run(search, search_dir)
     assert "--threshold-direction" in commands[0] and "less" in commands[0]
     assert "--block-size" in commands[0] and "42" in commands[0]
     assert "--max-rows" in commands[0] and "7" in commands[0]
-    assert "--query-smiles" in commands[1] and "CCO" in commands[1]
+    assert len(commands) == 1
+    assert result.metrics == {
+        "scanned_rows": 2, "valid_molecules": 2, "invalid_smiles": 0, "matches_emitted": 1,
+    }
 
 
 def test_runner_accepts_coordinator_workload_names(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
-    commands: list[list[str]] = []
-
-    def fake_run(command: list[str], **_: object) -> None:
-        commands.append(command)
-        output = Path(command[command.index("--output") + 1])
-        output.parent.mkdir(parents=True, exist_ok=True)
-        output.write_text("a,b\\n", encoding="utf-8")
-
-    monkeypatch.setattr("scimesh.worker.runners.subprocess.run", fake_run)
+    task_dir = tmp_path / "search"
+    task_dir.mkdir()
+    (task_dir / "input").write_text(
+        "chembl_id\tcanonical_smiles\nA\tCCO\nB\tCCCO\n", encoding="utf-8"
+    )
     task = ClaimedTask(
         "search", 1, "2026-07-30T00:00:00Z", "similarity_search",
         InputArtifact("https://example/input", "a" * 64), {"query_smiles": "CCO"},
     )
-    SciMeshRunner().run(task, tmp_path / "search")
-    assert commands[0][3] == "similarity-search"
+    result = SciMeshRunner().run(task, task_dir)
+    assert result.metrics["matches_emitted"] == 1
+    assert (task_dir / "result.csv").is_file()
 
 
 def test_claimed_task_rejects_path_traversal_and_invalid_metadata() -> None:
@@ -457,21 +548,15 @@ def test_relative_work_dir_is_normalized_for_runner_subprocesses(
 
     task_dir = config.work_dir / "task" / "1"
     task_dir.mkdir(parents=True)
-    (task_dir / "input").write_text("fixture", encoding="utf-8")
-    command: list[str] = []
-
-    def fake_run(args: list[str], **_: object) -> None:
-        command.extend(args)
-        output = Path(args[args.index("--output") + 1])
-        output.write_text("id,score\n", encoding="utf-8")
-
-    monkeypatch.setattr("scimesh.worker.runners.subprocess.run", fake_run)
+    (task_dir / "input").write_text(
+        "chembl_id\tcanonical_smiles\nA\tCCO\nB\tCCCO\n", encoding="utf-8"
+    )
     task = ClaimedTask(
         "task", 1, "2026-07-30T00:00:00Z", "similarity-search",
         InputArtifact("https://example.test/input", "a" * 64), {"query_smiles": "CCO"},
     )
     SciMeshRunner().run(task, task_dir)
-    assert command[4] == str(task_dir / "input")
+    assert (task_dir / "result.csv").is_file()
 
 
 def test_worker_registration_sets_returned_identity(tmp_path: Path) -> None: