feat(mcp): stdio MCP server with research tool

U3 of the Claude Desktop .mcpb bundle plan.

- internal/engine/run.go invokes python3 with the cached last30days.py,
  forwards os.Environ() so MCPB user_config env-injection reaches the
  engine, sets PYTHONPATH so the lib/ imports resolve, and surfaces
  three distinct error shapes (missing interpreter with install URL,
  non-zero exit with stderr, timeout). RunOptions.PythonPath lets tests
  inject a stub without manipulating PATH.
- internal/engine/run_test.go drives a shell-script stub interpreter
  through happy path, env forwarding, PYTHONPATH, non-zero exit with
  stderr surfacing, timeout, missing python3 (empty PATH), missing
  last30days.py, empty CacheDir, and timeout-env-override parsing.
- internal/tools/research.go registers a single research tool whose
  schema mirrors /last30days <topic> (required topic, optional emit
  enum, optional save bool). Validation failures surface as MCP
  tool errors so Claude sees structured failures instead of transport
  faults; engine extract or run errors fold engine stderr into the
  message so users can diagnose without leaving Desktop.
- internal/tools/research_test.go covers requireString, emitArgument,
  boolArgument, handler-level validation routing, and formatRunError.
- cmd/last30days-pp-mcp/main.go wires NewMCPServer + tools.Register +
  ServeStdio. main.Version is ldflags-stamped at build time and
  namespaces the per-user cache. mcp-go pinned at v0.54.0.

Smoke check: ./build/last30days-pp-mcp answers tools/list with the
research tool plus full schema and read-only/open-world annotations.
go test ./... passes across engine + tools (32 cases).

mcp/internal/tools/research_test.go

@@ -0,0 +1,145 @@
+package tools
+
+import (
+	"context"
+	"errors"
+	"strings"
+	"testing"
+
+	mcplib "github.com/mark3labs/mcp-go/mcp"
+
+	"github.com/mvanhorn/last30days-skill/mcp/internal/engine"
+)
+
+func newCallToolRequest(args map[string]any) mcplib.CallToolRequest {
+	var req mcplib.CallToolRequest
+	req.Params.Arguments = args
+	return req
+}
+
+// resultText pulls text content out of a tool result so tests can assert on
+// the body Claude will see. Returns empty string when the result is nil or
+// has no text content.
+func resultText(res *mcplib.CallToolResult) string {
+	if res == nil {
+		return ""
+	}
+	var out strings.Builder
+	for _, item := range res.Content {
+		if tc, ok := item.(mcplib.TextContent); ok {
+			out.WriteString(tc.Text)
+		}
+	}
+	return out.String()
+}
+
+func TestRequireStringRejectsMissingAndBlank(t *testing.T) {
+	if _, err := requireString(map[string]any{}, "topic"); err == nil {
+		t.Fatal("expected error for missing topic")
+	}
+	if _, err := requireString(map[string]any{"topic": ""}, "topic"); err == nil {
+		t.Fatal("expected error for empty topic")
+	}
+	if _, err := requireString(map[string]any{"topic": "   "}, "topic"); err == nil {
+		t.Fatal("expected error for whitespace-only topic")
+	}
+	if _, err := requireString(map[string]any{"topic": 42}, "topic"); err == nil {
+		t.Fatal("expected error for non-string topic")
+	}
+	v, err := requireString(map[string]any{"topic": "OpenAI"}, "topic")
+	if err != nil || v != "OpenAI" {
+		t.Fatalf("requireString ok = %q, %v", v, err)
+	}
+}
+
+func TestEmitArgumentDefaultsAndValidates(t *testing.T) {
+	cases := []struct {
+		name    string
+		args    map[string]any
+		want    string
+		wantErr bool
+	}{
+		{"missing defaults to compact", map[string]any{}, "compact", false},
+		{"empty string defaults to compact", map[string]any{"emit": ""}, "compact", false},
+		{"compact passes through", map[string]any{"emit": "compact"}, "compact", false},
+		{"html passes through", map[string]any{"emit": "html"}, "html", false},
+		{"invalid value rejected", map[string]any{"emit": "json"}, "", true},
+		{"non-string rejected", map[string]any{"emit": 7}, "", true},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			got, err := emitArgument(tc.args)
+			if (err != nil) != tc.wantErr {
+				t.Fatalf("err = %v, wantErr = %v", err, tc.wantErr)
+			}
+			if got != tc.want {
+				t.Fatalf("got %q, want %q", got, tc.want)
+			}
+		})
+	}
+}
+
+func TestBoolArgument(t *testing.T) {
+	v, err := boolArgument(map[string]any{}, "save")
+	if err != nil || v {
+		t.Fatalf("missing: %v, %v", v, err)
+	}
+	v, err = boolArgument(map[string]any{"save": true}, "save")
+	if err != nil || !v {
+		t.Fatalf("true: %v, %v", v, err)
+	}
+	v, err = boolArgument(map[string]any{"save": false}, "save")
+	if err != nil || v {
+		t.Fatalf("false: %v, %v", v, err)
+	}
+	if _, err := boolArgument(map[string]any{"save": "true"}, "save"); err == nil {
+		t.Fatal("expected error for string value")
+	}
+}
+
+func TestResearchHandlerValidationErrorsAreToolErrors(t *testing.T) {
+	// Validation failures are returned as MCP tool errors (not Go errors)
+	// so Claude sees a structured failure with a readable message rather
+	// than a transport-level fault.
+	handler := makeResearchHandler(Config{Version: "test"})
+
+	cases := []struct {
+		name    string
+		args    map[string]any
+		wantSub string
+	}{
+		{"missing topic", map[string]any{}, "topic is required"},
+		{"blank topic", map[string]any{"topic": "   "}, "non-empty string"},
+		{"invalid emit", map[string]any{"topic": "OpenAI", "emit": "json"}, "must be 'compact' or 'html'"},
+		{"non-bool save", map[string]any{"topic": "OpenAI", "save": "yes"}, "save must be a boolean"},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			res, err := handler(context.Background(), newCallToolRequest(tc.args))
+			if err != nil {
+				t.Fatalf("handler should not return Go error for validation; got %v", err)
+			}
+			if res == nil || !res.IsError {
+				t.Fatalf("expected IsError result, got %+v", res)
+			}
+			if !strings.Contains(resultText(res), tc.wantSub) {
+				t.Fatalf("result text %q missing substring %q", resultText(res), tc.wantSub)
+			}
+		})
+	}
+}
+
+func TestFormatRunErrorIncludesStderr(t *testing.T) {
+	res := &engine.RunResult{Stderr: []byte("engine exploded\n")}
+	msg := formatRunError(errors.New("boom"), res)
+	if !strings.Contains(msg, "boom") || !strings.Contains(msg, "engine exploded") {
+		t.Fatalf("formatRunError missed pieces: %q", msg)
+	}
+}
+
+func TestFormatRunErrorHandlesNilResult(t *testing.T) {
+	msg := formatRunError(errors.New("boom"), nil)
+	if msg != "boom" {
+		t.Fatalf("nil result: got %q, want %q", msg, "boom")
+	}
+}